To solve the location shooting problem of the elemental image array in integral imaging which is as the result of the pickup device limitations,a novel method to generate elemental image array from the sparse viewpoint images by parallel mapping was proposed.First,a camera array was used to capture the sparse viewpoint images of the objects.Then,the horizontal and vertical disparity maps of every image were calculated,respectively,and the spatial location of the actual object point corresponding to every pixel in the image was reconstructed.Finally,the elemental image array was generated by using the parallel mapping algorithm.For the holes which still exist in the elemental image,the interpolation algorithm was uesd to fill them.The captured sparse viewpoint images and the synthesized elemental image array were presented in the experimental results.The experiment results show that the synthesized image has continuous visual angles and can truly represent the space structure of the subjects.And the proposed method is superior to the traditional elemental image array pickup method in the implementation.

The phase part of short-time Fourier transform involved in most speaker recognition algorithms was often ignored because of the difficulties in processing the function expressions.The feature extraction procedure of interframe difference phase spectrum sticks to making a reasonable compromise between the two factors of the detection ability of sudden phase changes and phase warping,but the determination of frame length and frame shift based on empirical analysis lacks the rational basis.Thus the optimal setting method using mutual information criterion was proposed.The speaker recognition test experiments on TIMIT speech database show that the performance on the condition of the optimized frame length and frame shift is superior to that with empirical analysis.

Three methods of three-dimensional reconstruction technique were aralyzed and discussed.First,the process of Simulated Annealing(SA),the Fast Fourier Transform(FFT),Muti-Point Statistics(MPS) were introduced.Then,the differences of the three methods were analysized.Finally,according to pore parameter and characteristic functions,the results of various reconstruction algorithms with the real rock structure were compared, and the final analyzing results were given.

Aiming at the difficulty of achieving interactive simulation between soft tissue and suture, a method to simulate suturing procedure in virtual surgery was proposed.Firstly,mass-spring model was used as the physical model for soft tissue.In order to perform the relative motion between two masses of soft tissue realistically,an internal friction coefficient was introduced first.Secondly,the interaction of suture and soft tissue was aralyzed.A simple linear mass-spring model with following the leader algorithm was used to model the suture.A stress aralysis was carried out,when the suture touch the soft tissue,so the interaction process was close to physical reality.The movement of the whole suture computed by following the leader algorithm gave good performance in suture's rigidity characteristics.Finally,the interaction of the two deformable objects in a continuous pattern suture was considered,for each node of the soft tissue model,the node that attached to it was stored.This information can be used to compute the soft tissue deformation well.Experiment shows that it is realistical to simulate the soft tissue deformation and suture.meanwhile,the simulation is in real time.

Due to the information of computer-generated hologram is too enormous to transmit and store,a hologram compression algorithm was proposed based on artificial neural network.A set of training sample made of multi-dimensional vectors was obtained by preprocessing and used to train the neural network.The neuron numbers in input layer and hidden layer was adjusted so that the data volume of computer-generated hologram can be compressed with various ratios in both layers.Experimental results demonstrate that the proposed algorithm is uncomplicated and rapid.In addition,it has better self-adaptation and intelligent.The reconstructed image is obtained with relative large compression ratio and the quality is satisfied.

To solve the problem of losing details with traditional image fusion algorithms based on wavelet transform,a new CT/MRI medical image fusion algorithm based on wavelet domain was proposed.Average gradient and variance were used to determine the low-frequency component of the fused image;Gradient energy ratio weighted fusion rule was used to fuse the high-frequency components.Compared with the traditional methods of wavelet transform,the experimental results show that the new algorithm can fully fuse the two different modes information together and preserve the important features of the original images.The fused image contains richer and more comprehensive details,which increases the amount of the medical image fusion information effectively.

To solve the problem of serious cache conflict caused by traditional cache index mapping schemes, a new XY-type cache index mapping scheme based on the calculation from XY coordinates was proposed.A perfect scattering distribution of GPU cache index on frame buffer could be obtained,the bad effect caused by frame resolution was avoided completely.The evaluation results show that this scheme can reduce cache miss ratio by 82% at most.Since the direct-mapped or 2-way set-associative cache implementation can achieve the cache miss ratio close to full-associative cache,this scheme can decrease the design complexity and consumption power.

Due to the relative spatial complexity of mesh structure,the classic RayCasting method costs much time,which can hardly meet the real time requirement in most registration applications.A primitive-based RayCasting algorithm was proposed,with improvement on ray/triangle intersection and range filtering.Experimental results shows the efficiency of the method and it can also be paralleled.

To realize the accurate detection of facial features,a YUV color space and difference projection theory combined method was presented.The histogram curve fitting method was used to identify the location of the face;in the YUV color space,lip regions was got from the projection of images according to its red chrominance;vertical and horizontal difference projection theory was carried out to detect the eye regions and it was the foundation of facial expression recognition.Experimental results show that the method presented could realize the feature detection effectively,regardless of light changes.They also show that the algorithm is of high precision,robust and easy to use.

Pose estimation of different human faces is a key technology of varied human face recognitions which has not been well resolved yet.In order to get the mapping between face poses and feature spaces,an algorithm based on PCA was presented to realize the face pose estimations.Firstly,the subspaces and eigenvectors of different poses was built and selected by using PCA theory.Secondly,the Euclidean distance classifier was chosen to estimate the multi-poses of human faces.Lastly,the relationship between PCA's Variance contribution rates and pose estimation accuracy rates was investigated in detail.The algorithm was validated by means of experimental tests of human face pose database,which contains 23 kinds of poses and 630 samples in detail;experimental results show that the precision of this algorithm is above 84%.It also shows that the PCA based pose estimation method is feasible and valid for varied human face recognition.

An improved automatic synthetic registration method for substation multi-modal images is proposed.Two multi-modal images were different sizes,so wavelet transform was done.And the results were prepared for registration.In order to search the best registration position of two multi-modal images,genetic algorithm was used in the method.And one of the six methods was the fitness function of genetic algorithm in each search.Then six results were obtained for original images,and were integrated based on dynamic credibility.It is proved that automatic registration for multi-modal images of different sizes can be realized.This method is high speed and robustness,and it is more accurate than any single registration method.Besides,the credibility can be corrected dynamically.It has strong adaptability.

A video object extraction method based on active learning SVM was presented.By using the idea of active learning,the traditional SVM was improved,so the extracted video object was more accurately.An adaptive change detection method was used for obtaining the original video object,then the object was chosen for SVM training,and the positive samples were trained for constructing the intensifying classification plane.Experiment results show that the proposed method can overcome the shortage of the traditional SVM,get more exactly object edge,and reduce the computational complexity.

Lp(0<p<1) problem in sparse image reconstruction is a non-convex optimization problem.The non-convex Lp problem has been splitted into two sub-problems:X sub-problem and Y sub-problem by a half-quadratic penalty method.The closed form solution for X sub-problem could be obtained by the derivative of smooth function.And the solution of Y sub-problem was solved via shrinkage fixed point iterative formula.The splitting algorithm to solve Lp problem for sparse image reconstruction in compressive sensing was estabished by the alternating minimization method for the two sub-problems.Some different kinds of MR images was employed to test in the numerical experiment,and the results demonstrate that the non-convex splitting method is not only more accuracy but also lower sampling rate than convex splitting method in the sparse image reconstruction.

A new texture image retrieval method based on dual tree complex wavelet transform (DT-CWT) and Generalized Gaussian Density was proposed.By using the DT-CWT the query images and target images are decomposed to six directional sub-bands at each level.Modeling the marginal distribution of dual tree complex wavelet coefficients using Generalized Gaussian Density (GGD) to generate texture feature vectors. Kullback-Leibler distance (KLD) function was used as similarity measurement.The experimental results show that this method has higher accuracy than the methods based on energy feature and Euclidean distance as well as methods based on wavelet transform,contourlet transform and others using generalized Gaussian model in the same scale.

By using the CAAD system based on virtual reality,the designer should cost much time on modeling to architecture exterior environment,and the virtual architecture exterior environment was not quite true to the original CAAD system based on the technology of augmented reality was proposed.The 3D model of architectural design was registered in real world,and was integrated with architecture exterior environment by attitude tracking of real object and blending between virtual object and real object.Experimental results prove the possibility in technique.It also shows that this method designer can avoid modeling to architecture exterior environment and at the same time keep the architecture exterior environment in the simulation real.It is helpful for a designer to make reasonable assessment to architectural design.

Neural network ensemble (NNE) was applied to the pattern recognition of human arm motion with sEMG signal's analyzing.The feature vectors of sEMG were extracted with wavelet packet decomposition,then a NNE model was generated using Bagging algorithm and BP neural network was used as sub neural network.The output of NNE was achieved by relative majority voting decision method.Finally,recognition experiments were done with sEMG signals gathered from four different hand movements,and the results reveal that NNE can increase the correct recognition rates significantly comparing with single neural network,proving the validity and feasibility of using NNE in the field of human arm motion recognition.

It is almost impossible to achieve interactive rendering rates on a PC without hardware speedup when 3D visualization display technology is used in the field of medical image.A parallel rendering method of computing with CUDA applied on a graphics card was proposed.The MFC DLL and export class technique was used to avoid a mass of code migrating work.The most efficient number of parallel threads was also determined by analyzing the result data.The experiment shows that the speed of volume rendering based on ray-casting is improved by about 5～6 times

An efficient corrupted image (missing data) registration algorithm was proposed.Firstly,the lost elements in the corrupted image was recovered by using matrix completion,then PCA (Principal Component Analysis) was applied in SIFT (Scale Invariant Feature Transform).To solve the shortcomings of 128 element eigenvectors' large storage and more time to match,PCA was used to decrease the element of eigenvectors presenting descriptors.Finally,Gaussian weighted Euclidean distance instead of normal Euclidean distance was used to judge the similarity of feature points.The experimental results show that this algorithm, when applied in the computer vision field of content-based image and video retrieval,has good stability and accuracy to corrupted image registration.

In order to make the LED panel more prospective for parallax barrier autostereoscopic display and achieve virtual vision on the autostereoscopic display of LED panel.A special pixel arrangement of LED panel was proposed.To some extent the resolution of the LED autosereoscopic display was improved;A new way to calculate the viewing zone for the LED autosereoscopic display was proposed,base on LCD autosereoscopic display,and it can make the viewing zone of the LED autosereoscopic display more accurately.

Aiming at the severe spectrum distortion of the traditional PCA fusion method,starting from anlysing the sensor non-consistency of spectrum response range,different fusion strategies were used for distortion region and non-distortion region by pre-judging the spectrum distortion region of the fusion image.In the non-distortion region,the traditional PCA method was used to fuse. While in the distortion region,the first principal component was replaced by the wavelet fusion of it and the pan chromatic image.And a new multilayer fusion flow was propsed to improve the distortion problem of fusion image.Experimental results show that the proposed method can achieve high fidelity of the image fusion have hign adaptability.

A deformation detection method of interferometric synthetic aperture radar system which is called two tracks differential method was proposed.Unwrapped phase or phase gradient information to confirm deformation area were used in the proposed method.Detection Principle was described and detection flow was designed.Experiment results indicate that two track differential methods could detect deformation area.Comparative analysis shows that it has low sensitivity when unwrapped phase information is used to detect position and size of deformation area;and it has high sensitivity when phase gradient information is used to detect the position of deformation area only.

For the enormous data information contained in multi-view video,taking into account the difficulties of storage and transmission,the transform domain features of Fourier transform (FT) and multi-dimensional vector matrix discrete cosine transform (MVM-DCT) of multi-view video are deduced for the first time.Combined with the advantages that multi-dimensional vector matrix could put multi-view video into a multi-dimensional mathematical model,as well as orthogonality of DCT,MVM-DCT deduced could remove correlation among space,inter-frame and inter-view.The deduced formula shows that coefficient energy in high-dimensional transformation domain is mainly concentrated in a folding cuboid which attenuates along both the time dimension and view dimension.Then the experimental simulation verified their correctness.The significance of formula derivation is that they reveal the relationship between motion vector,angle variable and coefficients concentration in high-dimensional transformation domain.Finally,introduce minimize additional criteria about motion analysis briefly.All these will provide a theoretical guidance for the further compression of multi-view video coding system.

A series of 2D and 3D images' electroencephalography(EEG) using ERP technology were compared.By analyzing EEG signal of 2D and 3D image,it was found that amplitude of P300 component evoked by 3D images is higher than that by 2D images.Especially,in the unattended condition,the N140 component evoked by 3D images has a long latency.Paying attention to stimulus will evoke more EEG energy than that of paying no attention.

A new approach of image restoration for the sequence of fluorescein angiography was proposed.First,the intensity constraint was incorporated into Miller regularization equation to obtain the ability to preserve the high intensity of pixels.Thus,the restored image would be influenced not only by the smoothness constraints,but also by the intensity constraints.Secondly,in the process of the restoration,we use an intensity template obtained from the pre-filtering procedure was used to achieve the intensity constraints.The template represented an image composed by the desired intensity value.The experiments show that the proposed scheme gains a better result in both high intensity preservation and image restoration.

A method of expression capture of facial key points in the field of animation production was proposed.It only used an ordinary color camera for video capture,used the methods of image processing and pattern recognition on each frame to identify the actor facial key points (eyebrows,eyes,mouth,etc.) and their physical coordinates in the image.Associated with the animated character,facial expressions of the actor could be embodied with the character real-time.With the characteristics of simple equipment and high real-time,the efficiency of animation production is improved significantly.

The well-known Chan-Vese model has been widely used in image segmentation.However,the original model is limited by two important numerical issues.Firstly,the level set method does not implicitly preserve the level set function as a signed distance function.Secondly,the level set method is slow because of the gradient descent equation.In this paper,two fast algorithms,a dual method and a split Bregman method,were proposed to improve the computation efficiency.In order to force the level set function to be a signed distance function during evolution,a projection approach was proposed to solve the constraint,and then the augmented Lagrangian method was used to speed up the convergence rate.The experimental results demonstrate that the proposed methods not only have better performance,but also are more efficient than the gradient descent method.

A algorithm of shadow elimination based "Double rectangles" was optimized,it was in HSV color space.The algorithm may extract the moving object totally by reducing the influence of noise and has good real-time performance.It was realized on TMS320DM642 Video/Imaging DSP.The mode of video capturing was set and the program of DSP was optimized.First,the color space conversion was carried out.Secondly,rectangular moving region was first detected roughly according to the difference video frames based on V component and further accurate located by background subtraction video frames of S component.Finally,it fused targets information of each component in the detected moving region.The capture of targets was accomplished and it had good real-time performance.The results of experiments show that the motion detection algorithm can eliminate shadow fairly good and has good detection result of moving objects.

The theory of One Class Support Vector Machine (OCSVM) has an advantage over limited sample,high-dimensional space and unbalanced datasets.OCSVM parameter selection algorithm was improved by using a weight value simulated annealing method and dynamic inertia factor particle swam algorithm,as a result,the traffic classification accuracy is improved by nearly ten percent.Drawbacks such as low accuracy of the traditional traffic classification and overhead are solved.It is of great significance to improve the quality of network services, network management and control network security and so on.

The design and implementation of an image retrieval system integrated multi-features through combining color and shape were realized in this paper.In the respect of color feature,some improvements were made to the traditional color histogram method,and a new color histogram method based on main colors was proposed.By combining the major color retrieving method and the color histogram computing,two quick screenings were carried out,thereby the scope of the search was narrowed and the retrieval efficiency was improved.With Fourier shape descriptors adopted,an improved contour-based description method was proposed.Because the tangential angle of contours (curvature) was highlighted and factors such as complex coordinates and center distance were ignored,within a reasonable range,the accuracy was lowered appropriately and the query speed was improved significantly.Finally,the fusion of the HSV feature and ART feature,improved the retrieval accuracy of road traffic signs.Autonomous mobile robot platform was realized in VC 6.0.Experiments show that,if combining the features of the color and shape together,adjusting the weighting coefficient λ between the two,ultimately the value of λ can be determined,the recall ratio and precision ratio reach a relatively higher level.

According to the problem which image edge detection is easily affected by illumination,an edge detection method that using phase congruency in Riesz transformation space was proposed.Firstly,Riesz transformation instead of Hilbert transformation was used for image processing,new transformation space which is used for computing phase congruence is defined by different scale factors of Riesz transformation.Feature image was obtained by (computed).The feature of phase congruency algorithm in Riesz transformation space,and then Non-maxima suppression was uesd to detect edge information.Finally,compared with the traditional method of Hilbert phase congruence,2D Log-Gabor transformation and Canny edge detecting,the proposed algorithm can achieve promising results and be applied to images edge extraction under asymmetry quality of illumination conditions.Moreover,simulation results show that image feature extraction based on Riesz transformation operating is faster than the traditional phase congruency based on Hilbert transformation.The proposed method is feasible and valid.

In order to speed up the process of the orthogonal test for determining structure parameters of piezoceramics composite plate.Based on the Monte Carlo methods of reliability analysis,the determination of parameters of piezoceramics composite under the bending vibration model in the piezoelectric nozzles-flapper structure was studied.The probability design system in the ANSYS software was applied to calculate the impact of the non-deterministic input parameters on the structural properties.The ANSYS analysis results show that the largest sensitivity on the nozzle-flapper structure is the radius R_-1 of piezoelectric ceramic and then followed by the copper electrode thickness T_-2.The both impacts are negative.It indicates that they have the inverse relation to the structural properties of the nozzle-flapper.While the influence of the thickness T_-1 of the piezoelectric ceramic is positive.It indicates that it has the proportional relation to the nozzle-flapper performance.And the quantitative relationship among the three parameters is R_-1>T_-2>T_-1.The study results provide a basis for speeding up the process of the orthogonal test and improving test accuracy.

Hyperspectral remote sensing is the multi-dimensional information obtaining technology,which combines target detection and spectral imaging technology together.That is,it could obtain the two-dimensional object distribution information and one-dimensional spectral feature characteristic information at the same time.Compare with multi-spectral remote sensing,hyperspectral images contain abundant spectral information for the targets,which could greatly reflect the detailed characteristic of the ground,and makes the precise classification possible.Through the comparison of the theory between support vector machine(SVM)and relevance vector machine (RVM),this study utilized the two high-dimensional data processing methods in the classification of the same hyperspectral imagery.The experiment results show that the overall classification of SVM is slightly higher than that of RVM.

Because conventional anomaly detection algorithms are based on special assumptions,a new algorithm of hyperspectral imagery anomaly detection was presented.Without assuming the background model,first,iterative error analysis (IEA) was used for endmember extraction.Then the Spectral Similarity Scale (SSS) was measured.Through computing the kernel spectral angel cosine (KSAC),the anomaly detection result was obtained.The simulation result shows that the new algorithm can detect the anomalies exactly,and what's more,the new algorithm has the advantages of little computation time and high efficiency.

Four dimensional super plural structure of a quaternion is an orthogonal structure,which can keep the inherent orthogonality of each vector sensor component, and improve anti-interference ability and discrimination of the vector sensor array.Quaternion theory was applied into parameter estimation of uniform circular vector of array signals,the electromagnetic vector sensor array receiving signal model was established based on quaternion,and joint estimation of the electromagnetic vector sensor array signal's DOA and polarization information using quaternion's multidimensional orthogonal properties and the matrix theory combining with existing algorithm were realized.Simulation results show the effectiveness of the method.Compared with V-MUSIC,the quaternion model can obviously increase the signal parameter precision.

In order to predict and evaluate braking capacity comprehensively of designed products,CFD numerical simulation method was taken to predict dynamic braking performance of hydrodynamic retarder.Through simulating and analyzing,two important curves are gotten.One is the relation between braking moment and filling rate,the other is the relation between braking moment and braking time.All these work will provide more reliable method for predicting its capacity during the hydrodynamic retarder design.

A mathematical model using the spline functional as smooth constraints was presented.The second-and fourth-order partial differential equations constraints were two special cases in the model.The necessary condition for optical flow minimization problem solution was also presented.This model provided a basis for formal representation and numerical computation of optical flow from a methodological point of view.The significance of this mathematical model lay in the simplification of the equations for optical flow computation into linear algebraic equations.The simplification can contribute to discrete representation of the optical flow equation,and also verify that the use of smoothness constraints can ensure the existence and uniqueness of the solution from the view of the algebraic equations.

When target move fast or there were interference,the target tracking algorithm based on CamShift was often tracking failure only using color information.In order to solve this problem,an improved CamShift target tracking algorithm which combined with mobile target estimation was precented.By weakening color histogram information of the target in the background and estimating target motion,this algorithm still can realize effective tracking in the occlusion or interference of the tracked target.The validity of the algorithm is proved by experiment.

In order to reduce computational complexity and memory access store of H.264 fractional-pel interpolation,an improved filtering algorithm for fractional-pel interpolation was proposed, in which the filter coefficients are variable and the sum of all the coefficients is unchanged.Based on the fractional pixel interpolation algorithm for motion compensation in current video coding standard H.264,it determined filtering coefficients according to the variation of the motion vector.Compared to the algorithm in H.264,18% spatial complexity was reduced,computation complexity was about the same.The experimental results show that the PSNR is increased and the Bit-rate is decreased by the improved filtering algorithm and show the coding efficiency is enhanced.

In order to solve the problem that the ECG P,T-wave signal is complex,weak,difficult to identify as well as the problem of low efficiency and invalidity of the implementation of the identification algorithm,a combined algorithm was proposed using the fast speed characteristic of lifting wavelet analysis signal in time-domain and frequency domain based on the analysis to the principle of lifting wavelet algorithm.It combined the lifting wavelet transform with different OP,utilized the lifting wavelet to denoise,and indentified the P,T wave using composite algorithm in the reconstructed corresponding level low-frequency signal.A follower threshold function adapting to individual differences and abnormal ECG changes was proposed.Finally,the results of MIT-BIH ECG database experiments show that the proposed algorithm has high accuracy than traditional wavelet recognition method,at least doubling algorithm execution speed,and easier for hardware implementation.

By adopting linear quadratic gaussian (LQG) benchmark to assess engine idle speed control loop performance,LQG benchmark algorithms in two cases of control loop containing and without containing measurable disturbance were proposed.The benchmark obtained the optimal input and output variances under different weighting factor values by using impulse response matrices to solve optimal LQG control law.The trade-off curve and input and output variances indices can be provided.Performance assessment of the four PID controllers of the engine idle speed systems are analyzed base on the LQG benchmark.

As there are usually multiple hardware devices and multiple types of communication in hardware-based vehicle simulation systems,the management of the communication plays an important role.A new way to effectively manage different types of communication in hardware-based driving simulation system was proposed.By translating data streams of each device to messages of common format,the differences between types of devices and interfaces was concealed from upper layer of the program and it became possible for different types of devices and interfaces to be treated in an identical way.The proposed way of communication management has made the design work,the maintenance and extension of a simulation system easy.

Based on the representation of gait energy image (GEI),the method of texture analysis was used to describe the local space distribution of pixel brightness.Three different local textural features which are the range of local brightness (ROLB),the standard deviation of local brightness (SDOLB),and the entropy of local brightness (EOLB) were obtained.The classic Euclidean distance and nearest neighbor classifier were adopted to discriminate different patterns.The experiments on CASIA gait database demonstrate the effectiveness of the proposed method.

According to the characteristics of controlled model,the support vector machine system identification model was analyzed,and the lower limb rehabilitation model was studied by using the Matlab simulation modeling and the grid search algorithm was selected to train parameter to optimize the system,the simulation results show that the proposed method has advantage of accuracy and relevance.

According to the mismatch problem in less textured area of stereo match algorithms in computer vision field,in order to find the optimal disparity map of the stereo image pair,a method using the sub-gradient projection method to solve strict convex function which composed by the pixel error function of the shaped image pairs and quadratic smoothing factor of disparity map was proposed,which was restrained by the convex constraint set of total variation function.Experimental results show that the method can keep the edge of primal disparity map well,reduce the mismatch in the less textured area of disparity generated by primal stereo matching algorithm,and get the better optimization disparity map effectively.

There is a gap between low-level feature of image and high-level semantic understanding of users in the automatic image classification.The fuzzy-rough set theory was introduced into automatic image classification.During the mapping from low-level visual feature to high-level semantic feature,the problems of high-dimensional feature vector processing and the appropriate choice of descriptors in image classification processing were converted to fuzzy decision table,and the concept of the semantic proximity was used to verify the dependent relations of image attributes for attribute reduction,which could eliminate the redundant information and deduce the rule of image classification.In the end,the category of an image was determined by a membership degree function.An image classification system was developed and the accuracy of the classification results was 81.7%,experimental results show that the method has good accuracy and effectiveness,and can achieve a better image automatic classification.

Biological feature detection is a key issue for wireless body area network (WBAN).The correlation structure of the signal group and the cross-correlation structure were researched in this paper.The distributed compressed sensing was used.A new distributed coding reconstruction algorithm,Joint Sparse Model (JSM-1),was established.The model was applied to WBAN, and the joint sparse signal reconstruction algorithm was proposed.The performance of the proposed algorithm is evaluated through numerical results.

The video object extraction algorithm that is based on the multi-frame difference and HOS (High-Order Statistic),can be implemented through DSP(Digital Signal Processor).However,due to the large amount of the video image data,the real time of the DSP is not satisfactory.In order to improve the speed of the DSP,a sampling algorithm was presented.It took full use of the flexibility and effectiveness of the sampling algorithm.By running on the Blackfin-533 DSP,the feasibility and effectiveness of the algorithm is proved.

As a new distributed computing technology of mobile agent(MA),agent is completely exposed in remote host and is easy to be isolated and attacked by the malicious host.Homomorphism encryption protecting scheme was put forward,aiming to solve the security problems of mobile agents migrating across heterogeneous network execution environments.Homomorphism idea was applied to mobile cryptography,and encrypted mobile codes could be executed directly on different platforms without decryption.Encrypted results were decrypted by MA creator,and decreasing executing time of handling process effectively.Addition-multiplication homomorphism (AMH) encryption strategy was introduced,encrypting steps were used as follows:three address operands are encrypted by AMH,three address operating codes are encrypted by composite function,and the correctness and security of encryption and decryption method were verified in this scheme at the same time.Experimental results show that by contrast to data obfuscation method,nearly 1/3 executing time will be saved using this protection scheme and the malicious hosts are detected timely.

A new DVC algorithm for multi-view video coding was proposed in this paper.At the encoder,the blocks of original video frames were classified,and each frame was jointly decoded through finite iterations.Multi-view DVC side-information generation and CS reconstruction algorithm were integrated in the iterations.There were lots of advantages in CS-DVC framecork,such as low-complexity encoder,low bit-rate,and rate-control flexibility etc.This algorithm was consisted of independent encoder/decoder without any feed-back channels.This feature made the algorithm applicable to more extensive fields.The simulation results show that the proposed algorithm has low encoding complexity and can keep compression efficiehcy.

To solve the problems such as unreal effects,incomplete details in the rain scene,an algorithm to simulate water surface after rain hitting by combining Meatball with mass-spring water surface was presented.First,the raindrops model was estabished with particle system combining textures,and the water surface model was established with the mass-spring model.Secondly,the algorithm animated half globular bubbles after rain hitting with Meatball.Thirdly,water splash was simulated by means of the particle system and a gradual thinning and syncretizing technique.Finally,with the help of Cg language,the reflection and the refraction effect on the water surface,globular bubbles and water splash were simulated.The experiment shows that this approach can be used to simulate rain scenes efficiently and realistically.

In order to diagnose crop diseases and insect pests,computer vision technique was used to study three common insect pests of the fruit trees in the suburb of Beijing.These insect pests were dichocrocis punetiferalis,holotrichia oblita and marumba gaschkewitschi Bremeret Grey.It used computervision method to clean the background and get the pest's morphological characteristics such as shape parameter,eccentricity and round-like features.The discriminant function of the morphological characteristics was established,and its efficiency was tested.It is concluded that the morphological characteristics can describe the pest features and discriminate these three insect pests effectively.

A parallelized Annealed Particle Filter method via heterogeneous computing (P-APF) was proposed,to implement real-time marker-less motion tracking based on OpenCL framework.The overall computing procedure in P-APF was decomposed into several computational tasks with corresponding granularity.According to the degree of parallelism,the tasks were assigned to standard and attached processors respectively,to fully leverage heterogeneous computing ability.A novel task latency hidden strategy was proposed to further reduce time cost.Experiments on different human motion datasets demonstrate that P-APF can achieve real-time tracking performance without losing accuracy.With an average acceleration ratio of 106 compared to serial implementation,the time cost basically remains constant with the growth of particle number and view number in a limited range.

To solve the consistency problem of divided visual scenes,the neural oscillator network was established by the neural oscillation correlation theory in neural and brain science.The visual scene segmentation was implemented based on the neural mechanism, which was under the corresponding stimulation,the neurons became synchronized between same objects,and the neurons became synchronized between different objects.The experiment results by computer simulations show that the better visual scene segmentation is achieved by the system parameters adjusted.This method is superior to the traditional scene segmentation method in the implementation.

Based on the purpose of achieving the optimal balance point on the development costs of game action and facial expression capture system,operating results and the development efficiency by using the Kinect motion-sensory camera,3DSMAX and the open source OGRE graphics rendering engine were adopted to design and implement based on the game action and facial expressions capture system by combining Kinect motion-sensory camera,3DSMAX and OGRE.The system hardware is simple,cheap,and has the characteristics of low hardware requirements,high accuracy,intuitive real-time preview and alone operation.It can be widely used in the field of animation,film production and virtual reality.

In order to improve the rendering accuracy,an algorithm based on depth image and half pixel precision was proposed.Firstly the reference images were preprocessed,including the half-pixel processing and brightness correction,then forward 3D image warping is utilized to generate the image and depth map of the virtual view,inverse 3D image warping was used to fill part of holes caused by some factors,such as inaccurate depth information,and then ghosting was erased according to depth comparison,finally different methods was used to fill the remaining holes according to the types of the holes.The experimental results show that the proposed method can obtain satisfactory rendering results.

In order to improve the accuracy and efficiency of multi-objective extraction and recognition in robot soccer image processing,offline adaptive threshold value color classification was adopted,different colors of multiple targets were identified,which had certain flexibility to the color changes.An improved corrosion and expansion algorithm in mathematical morphology was proposed to filter the image whose color was classified,thus improved the image processing efficiency and enhanced the segmentation accuracy.Seed filling method was employed to simultaneously implement segmentation and multi-objective feature extraction,and thereby the times of pixel traverse was reduced and the efficiency of the algorithm was increased.It is illustrated by the experiments that the above method can make the robot soccer adapt to certain illumination change and make multi-objective segmentation accurately and rapidly.

As different illumination scenes influenced on the result of the algorithm,we proposed an eye states detection algorithm based on phase information.After using phase congruency to extract the feature of eye,the result image and calculate the number of white pixels was culated to examine the eye states according to vortical histogram.The experimental result shows that it is robust to different illumination environments and has good accuracy.

Following Itti model's idea of the saliency extractiong using central-surround center-surrounding difference,a visual salient region extraction method based on relative entropy difference was proposed.First,the difference among all the image pixels was calculated;then,the central-surround region histogram difference was analyzed;finally,the visual salient region was extracted using the relative entropy theory.This method can be efficiently implemented because of avoiding the highly complex Gabor filter extraction of Itti model,and the extracted salient regions are the same with the human visual system.

According to the background modeling thought in background subtraction algorithm and developed from the point of view of extracting scene knowledge to establish the scene knowledge base,an algorithm of moving object detection based on scene knowledge was proposed.In this algorithm,the features of each segmented regions were extracted to build a scene knowledge base after we used an improved mean-shift method to segment the scene image;The feature vectors from the new frame image were extracted,and the information of the moving objects were detected by matching the new feature vectors with the pre-existing feature vectors we can get from the scene knowledge base.The simulation result shows that the proposed method can detect the moving information of original object and new object in the scene effectively and is robust to the noise of shadow and the deformation of the object.

Based on the Brown's theory,an effective method for the image matching was presented.Firstly Gauss filter and Wallis filter for pre-processing were used,then feature points were extracted by a simplified SIFT algorithm,finally the exact matching of the images was achieved with a bidirectional matching algorithm.The experimental results on layout with defect show that the method is of great advantage of achieving more matching points,being more accurate and leaving behind no duplicate point.Finally,the method is used in automatical detection of net and the result is good,and for the layout optimization of foundation.

As videos such as bloody, violent and pornographic could not be updated frequently,a new objectionable videos detection method based on the combination of machine detection and manual intervention was proposed.First,objectionable video was labeled by manual method.Then the normalized energy curve was calculated through the multiple color space and the video character database was set up.Finally,the system can automatical detect the objectionable videos by querying the video character database.Experiments show the method can accurately and quickly realize the detection and early warning of the objectionable videos.

A real-time detection system of moving objects in complicated static background based on FPGA was designed,and the system hardware platform and the software algorithm process were introduced.Background difference algorithm was used to detect moving objects.Median filtering and pixel shift algorithms were used to remove the noise.Erosion and dilation and Sobel edge detection algorithms were realized.Experiments show the system has clean background and detect the objects accurately and alarm in time.The processing speed can meet the needs of intelligent video surveillance system in real time.

In order to understand the underlying causes of visual discomfort so that it can be minimized or even eliminated,it is very essential to evaluate the visual comfort of 3D displays.The visual comfort is highly correlated with the disparity information of 3D contents.Hence,a subjective evaluation experiment was carried out to measure visual comfort of stereoscopic video at different disparity levels.Results show that the degree of visual comfort decreases as disparity increases and the comfort disparity is in the range of -119 to +115 minute of arc.

The method based on backlight parallax barrier was proposed to reduce crosstalk and provide 3D image display with full resolution.The proposed display consisted of a backlight parallax barrier and a 2D display panel.The backlight parallax barrier was an OLED panel.By lighting pixels in particular areas,the OLED panel was easy to adjust the width and location of the transparent stripes in parallax barrier.Synthetic images from 2D display panel were modulated by the backlight parallax barrier,and displayed 3D images.The principle and parameter calculation were described.A prototype of the proposed display was developed.It shows that by adjusting the width and location of the transparent stripes in parallax barrier,we can reduce crosstalk.And the time-multiplex technology is helpful for full resolution display.

A parallax-controlled multi-view image rectification algorithm was introduced.Through rotating,vertical parallax adjustment and parallax balance,the information of multi-view image could be rectified.The stereo image could meet the requirements of multi-view stereoscopic display.The object could be displayed with plus,minus and zero parallax.Experiments show that the rectified stereo images can be displayed on the multi-view stereoscopic display with excellent stereo perception,and the target object can be outstanding,sunken or on the screen on demand.

Restricted by the size of the mobile device' screen,the definition of the stereo video declines significantly when the original multi-view stereoscopic video display method based on large screen is applied to mobile devices.A parallax adjustment method was proposed based on the research of the causes and solutions to the fuzzy phenomenon of stereo video on mobile devices,together with the relationship between parallax and stereoscopic video depth.This method improves the definition of multi-view video significantly at the proper reduction of parallax,which,makes the viewing effect more clear and comfortable,and is suitable for the mobile devices and other small screen applications.

Aiming at characteristic of mass data of image out from high-frame camera,a high-speed real-time image acquisition and processing system based on CameraLink interface was designed and implemented.The FPGA chipset EP3CF484 accomplished image acquisition and image pro-processing,and the high performance DSP chipset TMS320DM642 accomplished more complex image processing.The design idea,hardware architecture,and functionality of the system were introduced,and then the design method for each functional module of the system was described.The experimental results demonstrate that the system can reach 500 frames per second for real-time image acquisition and processing.At the same time,the system can complete the video interface conversion from CameraLink signal to SDI (Serial Digital Interface) signal.

With the high resolution SAR applied widely in the military field,camouflage effectiveness assessment for military targets against SAR is urgently needed.Using SAR images acquisited from the test,distribution characteristics of the camouflaged targets and background in high resolution SAR images were analysed.The theory of hypothesis testing was introduced,the statistical distribution characteristics were verified.A new camouflage effectiveness assessing method was given.The contrast results of the high resolution SAR image interpretation and results experimental of the algorithm prove the validity of the method.

A stereo matching based on image segmentation and variable windows is presented was proposed.Firstly,the reference and the target image was segmented based on the color information,according to the hypothesis of disparity in the same color segmentation region smooth,the matching cost of segmented regions was calculated,and then the matching cost of the best variable windows was calculated through the average matching errors,error variance and biases error,at last,two categories matching cost were merged and the dense disparity map was abtained through local optimization.Experimental results show that the algorithm can get a good effect on low texture and discontinuous depth region,not only achieves high accuracy but also has shorter matching time.

The basic principles of electronic image stabilization (EIS) technology was applied.The image sequence of airborne imaging equipment was taken as the research object.The feature matching by feature points was utilized to calculate the motion vectors between frames.Kalman filter was applied in the motion parameter estimation to remove the jitter component.To achieve motion vector compensation,the improved frame by frame compensation methods was presented.After contrast of the interframe difference in the video sequence between before and after stabilities,it is well done to remove the jitter of the video sequence.

In the three-dimensional display,in order to research the effect of overall image quality when quality of monocular image declined,a method was proposed to add crosstalk based on image processing without changing the device itself.With the help of Matlab,there were different crosstalk levels of images and different crosstalk types of images.The relationship between different crosstalk levels of images and different crosstalk types of images were analyzed based on the perception experiment.The experiment results verified the relationship between the monocular image quality and overall image quality.It can be concluded that there is only a marginal difference between monocular image quality and binocular image quality when the crosstalk level is low.When it comes to high crosstalk level,the visibility of the binocular crosstalk is higher than monocular crosstalk.

In a video surveillance network system constructed by multi-cameras,when a target access to one camera's FOV(Field of Vision) from another's,it was a key problem to schedule the cameras correctly and to achieve reliable target handoff.Cameras scheduling and target handoff strategies were proposed based on topological model and feature learning that can be applied to multi-PTZ cameras relay target tracking system in a large and complex scene.First,a scene map model was constructed for Cameras,and a method based on real time background estimation was used to determine the handover time and position.Then a multi-camera geometric topological model was created,and,a scheduling handoff algorithm based on topological model and a feature learning algorithm were proposed to schedule the handover camera quickly and to complete the target handoff.The experimental results show that the proposed method can judge and schedule handover camera quickly and accurately in multi-cameras active tracking.

Most current binocular vision measurement algorithms are based on PC as a platform, but it has deficiencies in real-time and portable.Based on two DSPs as the core,extraction and matching of image corner,binocular ranging and other aspects were studied and improved.Without affecting the accuracy of the entire system and aiming to achieve accurate and fast three-dimensional measurement,a quick solution to solve the non-contact measurement was provided.

Speckle reduction anisotropic diffusion (SRAD)is an important method to denoise speckle noise.A new method was presented to denoise optical coherence tomography (OCT) speckle by combining SRAD and dyadic wavelet.Interscale dependency of dyadic wavelet coefficients was used to distinguish edges from noises.A quantitative estimation was provided to evaluate the probability that a pixel lied in edge-related regions or homogenous regions,and the probability was used to modify the diffusion coefficient in SRAD to get a new anisotropic diffusion method,which exhibited an increased ability to detect and suppress strong noises,while still equally was good at preserving image edges.Experiments show that the denoising performance of our method is better than that of SRAD.

Two dimensional all phase image interpolation template with standard diamond-shaped support was constructed.Using all phase linear filter theory and orthogonal U transform's sequence characteristics,we deduced all phase U sequence filter was deduced,and then image interpolation template was constructed.According to the constructed template,a new image interpolation algorithm was proposed.The experimental result shows that the proposed method is optimal compared with bicubic interpolation and the other image interpolation methods based on all phase filter.

A simple and practical method based on three-dimensional board was proposed for on-board calibration.The on-board camera was calibrated based on a three-dimensional calibration board,which was easy,high-precision,and convenient for follow-up testing and other works.The experimental results and the error analysis show the feasibility of proposed approach.

Image matching performance of normalized mutual information with different histogram bins was analyzed and discussed.Taking normalized mutual information with 16 histogram bins as similarity criteria,a fast image matching method was proposed based on an improved self-organizing hierarchical particle swarm optimizer with time-varying Acceleration Coefficients (IHPSO) through increasing the population of particles reinitialized and improving convergence criterion.Taking remotely sensed imageries captured by different sensors at different time as testing data,the algorithm with the exhaustive search method based on mutual information,Standard PSO and some improved PSO were compared.It is proved that the algorithm suggested has better matching performance and can be applied to the areas of fast image matching,e.g.aero craft navigation,positioning and movement tracking.

Based on super-resolution image reconstruction algorithm of sparse representation,this paper presents an improved use of Locality-constrained Linear Coding (LLC)super-resolution reconstruction algorithm that can improve the quality of the image reconstruction while reducing the computational complexity and improving processing speed,which achieves good results.First a pair of high and low resolution Dictionary was trained in accordance with a set of high-resolution images,then LLC method was used to encode the input image with the low resolution dictionary.A high-resolution image was preliminarily reconstructed with high-resolution dictionary and coding coefficient,and finally a global constraint was added to rebuild the high-resolution image.The experimenttal results show that the new algorithm, comparing to the criginal algorithm, not only improves the quality of image reconstruction, but also reduces the computational complexity.The above improved algorithm has also been extended and applied to multi-frame image super-resolution reconstruction. Analysis and experiments show that this super-resolution reconstruction algorithm achieves some good results, it can improve the quality of the image reconstruction while reduce the computational complexity with improved processing speed.

Now existing image feature matching algorithms are always high complexity and long time-consuming.A novel feature matching algorithm was proposed based on local feature points.Scale pyramid should be constructed first in which FAST key points were detected and extracted according to their Harris response.Then directions were distributed for key points using a method of intensity centroid.Finally,key point vectors were built via a sampling pattern.The hamming distance between the key point vectors in different images decided whether the two of them were matched or not.Experiments show that this algorithm is robust and reliable even under the condition of a certain degree of scaling,rotation and the effects of noise.Moreover,this algorithm is several times faster than SIFT while performing as well as SIFT in other aspects.

In order to reduce system delay,and meet the requirements of real-time,low time delay rate control algorithm for H.264/AVC.The algorithm was based on the characteristics of the H.264/AVC coding standard,fine layered rate control was performed and the forward rate control technology based on image complexity analysis and backward rate control based on feedback buffer capacity technology were integrated.Simulation results show that,compared with JVT-H017 algorithm,the algorithm can be effective in reducing the number of skipping frames to reduce delay and improves the timeliness of the video communication;better maintains the quality of image coding;PSNR dropped only by about 0.1 dB.The algorithm for computing the amount is small,and it is easy to implement.

In order to detect the snow coverage status such as mild or heavy coverage on road to keep car running safe,snow coverage detection method together with naive Bayesian classification algorithm was proposed.Firstly,road visual feature was extracted from the video by the method of machine vision and video object segmentation.Then the road snow coverage status was classified with naive Bayesian classifier.The performance of KNN,ANN and SVM classifier for road snow coverage problem were compared.

For vision navigation of ALV in complicated environments,a novel dictionary learning and sparse representation based road segmentation algorithm was proposed.The local image patch was used as the processing unit;a dictionary was learned based on man-selected typical road image and the dictionary could be updated in real time by online dictionary learning with the little piece of image right before the vehicle as supervision.With this dictionary,the on-road patches could be sparse represented precisely while the off-road patches could not A dictionary Learning and sparse representation based on classification framework was built and the local image patches could be classified by the reconstruction errors of sparse representation.A variety of experiments show that the proposed algorithm is suitable for various unstructured environments and is robust to illumination,shadow and water stains;unstructured environment.

Images captured in foggy weather condition exhibit losses in quality due to atmospheric particles.A novel single image fog removal method was proposed using improved atmospheric veil inpainting.At first,the atmospheric scattering model was simplified,and the coarser atmospheric veil estimate was refined using guided filter approach that improved atmospheric veil inpainting.Then a semi-inverse method of the fog image was produced to estimate the Airlight accurately.Finally,the fog-free image could be recovered by solve imaging physical equation of atmospheric scattering model.Results demonstrate the power of the proposed method performs as an edge-preserving smoothing operator and achieves a great improvement in image visibility.

Based on the optical characteristic of optical-plate and the structural characteristic of LED panel,a stereoimage synthesizing method based on LED panel had three steps:First,a pixel mapping which was used to synthesize the 3D image was calculated according to the parameters of LED panel and optical-plate.Secondly,multi-view images were processed by seam-carving based on the demand of the derived pixel mapping.At last,the 3D image was synthesized from the processed multi-view images.Experimental results show that the method,which gets better 3D visual,can be applied to any LED autostereoscopic displays,and can synthesize the 3D image of arbitrary viewpoints quickly and efficiently.

A multimedia terminal circuit based on DSP and FPGA was designed.The circuit function consisted of two channels of composite video input,one channel of selectable output of VGA and composite video,ethernet communication and stereo input and output.According to the SI theory and project requirement,rational layout and lining were made.The SI analysis including reflection and crosstalk was done with the high speed signal.The topological structure was adjusted and the circuit was optimized by using series resistor and modifying the layout.From experiments,the circuit has fully functions and reliable performance,and achieves the expected design requirements.

A new voltage-mode universal biquad filter circuit using current-feedback operational amplifiers (CFOA) was presented.The circuit consisted of four CFOA,two capacitors and ten resistors which achieved all the five filter functions,such as low-pass,band-pass,high-pass,notch,and all-pass from the same configuration.Circuit implementation was given and a computer simulation of the circuit was completed.PSPICE simulation results show that the design of circuit has the characteristics of simple structure and low sensitivity.All the capacitors are grounded and the filter is suitable for fully integration.

According to the problem of keeping image border and details weakly by MAP algorithm, a new method which mixed MAP and POCS for image super resolution based on Markov Network is introduced.The MAP operator is introduced to the POCS algorithm as a convex set of constraints.The POCS algorithm can be a good complementary to the MAP algorithm to overcome their original defect.Experiment shows that the proposed algorithm has better image restoration effect than the traditional algorithm.

Aiming at solving the problem that the time-frequency peak filtering(TFPF) hardly recover the subtle structures of the seismic events as attenuating random noise,different windowing lengths need to be set respectively for signal and noise.A fuzzy time-frequency peak filtering (FTFPF) algorithm was proposed using multiple time-windowing lengths.The new method used a fuzzy determination threshold to identify the membership of each sample to the noise segment and signal segment in seismic data.Then the multi-level windowing lengths were set for FTFPF according to the noise intensity in noise segments and frequency characteristics of signal in signal segments.The longer windowing length was used for noisy segments to remove random noise effectively,and the shorter windowing lengths were used in signal segments to preserve the details of the signal.The FTFPF was applied to synthetic records and common-shot point record,the model analysis indicates that the fuzzy discrimination can identify the signal and noise correctly.The filtered result by FTFPF has better ability to remove noise and preserve signal than that by TFPF.The signal-to-noise ratio (SNR) of the seismic image has been improved.

Traditional transition region extraction methods are based on gradient operator.The conventional gradient-based methods are sensitive to noise.According to the characteristics of transition region,the segmentation method based on information measure was presented.Then optimal segmentation threshold was attained according to the transition region histogram.The transition region was obtained by local entropy information measurement.The proposed method depend no more on L_low and L_high.The filtering ability of local entropy information measurement improves the ability of proposed method to deal with noises.Experimental results indicate that the proposed method outperforms the other methods.

Theoretical analysis and experimental study were applied in the research of the impact of network transmission errors on the multi-view 3D video quality.Error propagation of multi-view video over packet-loss networks was analysed as well as the theoretical formula of error propagation was derived.The impact of packet loss and bit error on the 3D video quality was studied though simulation methods.Acceptable loss rate and error rate of the stereoscopic video transmitted over the IP network were approximately determined according to the tolerance of the human eye on the decline in the image quality and depth sensation.

A neural network-based face detection system was proposed.Unlike similar systems which were limited to detecting upright frontal faces,faces were detected at any degree of rotation in the image plane.Multiple networks were employed;a "router" network first processed each input window to determine its orientation and then used this information to prepare the window for one or more "detector" networks.The training methods for both types of networks were presented.Sensitivity analysis on the networks were performed and empirical results were presented on a large test set.Finally,preliminary results were presented for detecting faces rotated out of the image plane, such as profiles and semi-profiles. Experimental results show that this system has better performance in accuracy and detection speed,and it is faster than a single network system.

Considering model dispersion and nonlinear in multimode fiber(MMF),MB-OFDM UWB technology was fused into MMF communication system.Through the analysis of a distinguishing feature of MB-OFDM UWB signal and model dispersion in MMF,it was proved that MB-OFDM UWB technology had the ability to overcome multimode fiber mode dispersion.Peak average power ratio(PAPR) of MB-OFDM UWB signal and the PAPR reduction method were researched.The simulation platform was established by MATLAB and Optisystem software,verifying the ability to overcome modal dispersion in MB-OFDM UWB MMF communication system and the improvement in system performance by the PAPR reduction method.

Although the robustness to noise of group delay function (GDF) was proved,the spiky effects caused by the resonance seriously undermined its further application.The ability to represent the original acoustic space was by no means guaranteed for the current improvement solutions.On the premise reducing the information loss in the process of feature extraction and aiming at decreasing its dynamic range,the feature of exponent function warping group delay function was produced.The test experiments of voice activity detection indicate that the robustness and detection accuracy is raised remarkably compared to the present GDF versions.

Based on the problem of color histogram,an improved color histogram extraction algorithm that based on the block and main color was proposed and an improved SIFT features extracted algorithm that based on the harris corner was proposed based on the problem of SIFT feature.Finally,a image retrieval algorithm that fused the improved color histogram extraction algorithm and improved SIFT feature extracted algorithm was proposed.The experiment shows that the algorithm is better on function and performance than any other algorithms,and it has good robustness.

A new of video dynamic watermarking algorithm based on DWT was proposed,parity frame of the video was separated from the video,the low frequency component after the decomposition of the secondary DWT of the parity frame was used to complete dynamic embedded watermarking frame,watermarking was recoveried by selecting integer coefficients to expand the selection of the watermarking information.Simulation results show that the algorithm can recover the watermarking better from embedded vodeo under resisting the attack of lossy compression and filtering.The concealment of the video is enhanced while the video robustness is maintained.

In order to improve classification performance of image,one efficient method was to combine different features together.The combination in various layers (feature layer and coding layer) was studied with the help of SIFT feature and normalized color histogram (NCH) feature,their effects were also compared.Experiments on several popular datasets show that,compared to combined features,the feature combination after feature extraction can reduce feature dimension while keep the similar classification rate,and the feature combination on coding layer can achieve even higher classification rate.It is concluded that feature combination should be better performed on higher semantic level.

The operation principle and parameter calculation of the 1DII displays using parallax barrier and lenticular lens array were described in detail.Two prototypes were developed.The experimental results show that the1DII display using parallax barrier is simpler and has lower cost,while the 1DII display using lenticular lens array provides brighter 3D images.

A coherent imaging system for large size high quality 3D display in the air employing low resolution spatial light modulators was introduced.It decomposed and reconstructed optical wave-fronts based on digital optical phase conjugation using an adiabatic taper consisting of a large number of single-mode waveguides.Due to reciprocity of an optical path it canceld automatically the aberrations inherent in traditional optical lens systems and provided very large field depth.Real 3D images could be formed in front of the screen,while virtual 3D images could be created behind the screen.All of the 3D images could be watched at the same time by many observers with naked eyes.As the calculation show the imaging resolution could vary from 1 to 112 micrometers when 3D images are formed at a distance from 0.1 to 10 m in front of the screen.

To solve the spatial information loss problem,the least squares support vector machine based on matrix patterns was applied to the recognition of lung nodule using image matrix as input.20 sets of CT images were used in the experiments.The 20 true nodules and 20 false ones extracted were used to test the function of classifier.The cross validation of the regularization parameters was obtained by the grid searching method to get the optimal parameters of the linear kernel function.The experimental results show that this method is feasible and effective for the detection of lung nodule.

Topology and direction were two important features of the object in space,and single spatial knowlege processing was unable to meet the demand of the theory and practical application.The interactive table between topolegical relation and basic cardinal direction was established based on the formal defination and compatibility isiues of basic cardinal direction, the compatible algorithm combined the constraints beween RCC8 and basic cardinal direction. The experimental results prove that the time complexity of the proposed method is O(n³).

During the high-frequency effective signal processing of seismic exploration records, the traditional Time-Frequency Peak Filtering (TFPF) may lead to wave distortion and large energy loss. In this paper, the Parallel Radial Trace Time-Frequency Peak Filtering (PRT-TFPF) arithmetic was adopted to process the seismic data. The proposed arithmetic has better performance to process the high-frequency effective signals with highly nonlinear degree by comparing with the TFPF method. The PRT-TFPF not only can decrease the signal frequency, stretch the wave form,but also can meet the unbiased estimation requirement in one data window lengthof TFPF, can reduce the energy loses of filtering process, which can take advantage of recovery of the signals. The results of the experiments on 60 Hz effective signal event show that the PRT-TFPF can remain the signal waveform without distortion, keep the amplitude, and recover the event continuously and clearly. By comparing the traditional TFPF, the Signal to Noise Ratio (SNR) of the proposed PRT-TFPF is improved by about 6.5 dB.

Through studying and simulating the traditional algorithm of image fusion and Inspiring by the multi-objective particle swarm optimization, we proposing an improved algorithm of MOPSO. The new algorithm based on multi-objective particle swarm algorithm framework. However, there are some differences between them. The new algorithm adopts more effective ways of speed changing and multi-objective choice processing which makes better performance and the searching solutions closing to the Pareto optimal solution set. The new algorithm has been used for remote sensing images fusion and multi-focus images fusion, which have achieved better results.

BP neural network prediction model was used in diseases prediction optimized by DS evidential reasoning in order to solve the low recognition rate and credibility of BP neural network processing the multi-target prediction. From the experiments on heart disease data set, the proposed model performs well in the accuracy of prediction and its robustness.

The fuzzy clustering method was proposed based on the transitive closure is and it was applied to web log clustering. The experiments show the algorithm is more effective in web user clustering and avoiding falling into the local optimal solution. It was compared with fuzzy C-means algorithm. The running time and error rate of the algorithm were discussed.

The robust representation of image is established by Histogram of Oriented Gradient (HOG). According to the processing mechanism of human visual system and the theory of multidimensional space biomimetic informatics, a biomimetic classification and recognition method of human behavior are proposed, which is based on HOG+SVM. The method is evaluated and compared with other commonly used methods. Results show that, for the classification and recognition of human behavior in still image, the proposed methods have better performance in recognizing different kinds of behavior, but the performance in recognizing similar behaviors still needs improvement.

In order to evaluate poultry freshness in a quick and non-destructive way, an intelligent inspection system was established based on neural network recognition technique and two types of sensors, including electronic nose and mechanical parameters. Sensors were chosen according to chicken deteriorate mechanism. RBF neural network was used as method of pattern recognition. The accuracy of the system for chicken freshness detection is up to 96.45% through physico-chemical examination. The experimental results indicate that it is feasible to integrate multiple sensors for chicken freshness detection.

For solving the sensitivity analysis of vibration systems, a rapid calculation of state vector sensitivity analysis derivatives was established. Based on the generalized modal theory, the sensitivity analysis of near defective systems could be transformed into one of the defective systems. A relaxation factor embedded in the combined approximations approach ensures the non-singularity of coefficient matrix. Moreover, the theory and method of near defective systems sensitivity analysis was given. The numerical example indicates the effectiveness of the algorithm.

FJC developed the Active seat Back Control Systems(ABCS) and did some improvement on seat structure to reduce the injury. After Euro NCAP Whiplash test, the conclusion is that using ABCS or improvement on the seat frame and foam can enhance Euro NCAP Whiplash test score and reduce passenger injury.

Model predictive controller, engine torque PI controller and braking torque fuzzy controller was designed based on the three degrees of freedom vehicle model. According to the simulation of single lane steering and brake typical conditions, the design for the vehicle stability controller has good control effect, can obviously improve the vehicle handling stability; The development of braking torque algorithm can make full use of each wheel braking force, makes the engine torque change and active braking pressure input smaller, obtains good handling stability of vehicles.

Vehicle seven degrees of freedom dynamic model was established,based on the fuzzy and PID the top controller of vehicle chassis integrated controller was designed, using PI method to control steering by wire system; using quadratic programming method for active yaw moment assigned and in the single lane steering typical condition was simulated, the results show that the design of the chassis integrated controller has good control effect, it can obviously improve the vehicle steering stability. The development of hierarchical control algorithm can make full use of various actuators, when the active steering Angle and active braking pressure input is smaller, it can still obtain good handling stability of vehicles.

This paper studied on the neural network modeling of X-by-wire vehicle ideal characteristics reference model to achieve "car adaption for driver", which are achieved by integrating the X-by-wire systems. The driver characteristics were classified and the control principle of car ideal characteristics was given. The car driving simulator experiment was designed and the neural network modeling for the ideal characteristics reference model was studied. The accuracy of model, degree of reflecting characteristics and real-time of model were verified by experiments. The results show that the reference model with ideal characteristics has high precision, real-time and can reflect different driving characteristics, so the method is feasible.

Estimation of blind zone of rearview mirror of most driver is one important element for design and development of BSD system or SWA system. Most present methods estimate the blind zone of drivers according to single eye position, which cannot reflect the blind zone of rearview mirror that is influenced by dynamic position of driver' eye and different eye position between drivers. A method of simulation of drivers' view from rearview mirror is proposed. Based on it, an eyellipse is used to locate drivers' eye position in order to estimate the probability of view of rearview mirror of most driver. Finally, the probability of blind zone is estimated. The proposed method provides a more accurate simulation on driver's blind zone, and present the differences between blind zone of different drivers.

In order to study cooling problem on the hydrodynamic retarder of heavy vehicle, the hydrodynamic retarder D300 is chosen. The hydrodynamic retarder D300 with liquid-filled ratios of 100% is numerically simulated by using the software of CFD. The closed temperature without heat exchanger is gotten in the condition of the same time and the different rotating speed. The plate-fin heat exchanger is chosen when matching heat exchanger, because it ensures that the hydrodynamic retarder can work at normal temperature. The serrated fin is chosen at the same time, in order to improve the cooling efficiency of the fin. According to the cooling needs of the hydrodynamic retarder D300, the heat exchanger that can meet the use of heavy vehicle is matched eventually.

To avoid the traditional trial-and-error methods generated during the manufacturing process of blades, this paper proposes to precisely control the forming of pump and turbine blades of stamping welded hydrodynamic torque converter YJC265, which contains stiffener and flanging. This method can predict the springback distribution accurately, compensate die surface of blades based on the calculation results and obtain ideal stampings. Furthermore, a YJC265 hydrodynamic torque converter prototype has been made through the forming and springback calculation and die compensation method, and the prototype performance test results exceed the design requirements, production cycle and cost are also far less than the traditional method. This new method can guarantee the springback accuracy of blades, guide die design, ensure a success rate of die development and is of great significance to the actual production of the torque converter blades.

According to the comprehensive functional asphalt pavement performance requirements, by GA-ANN optimize mineral aggregate ratio of asphalt pavement was optimized using GA-ANN. A seties of factors such as asphalt mixture type of mineral aggregate, asphalt type, and fill plus were comprenensively surveyed. Evaluation indicators integrated dynamic stability as elegant residual stability, creep rate, friction coefficient and permeability coefficient, and porosity and other factors, which put forward the best strategies for optimizing functional asphalt pavement mineral aggregate ratio. Comprehensive evaluation results show that the scheme is feasible and effective, and to some extent, to meet the requirements of the cause of chinese road transport functional asphalt pavement material ratio.

Based on ax-symmetric jet and plane jet ,this article stated an inference about air ventjet angle of automobile instrument panel from Prandtl's mixing length hypothesis and the theory of parallel multiplejets, which was confirmed by finite element simulation analysis and tests.The results of the study on vehicle engineering show that increasing the number of vanes doesn't change yet angle.

The modeling of ring model was designed by ANSYS, contact problem of gear pair was made to contact problem of ring model. According to finite element analysis method of contact problems in ANSYS and optimal design technology of bionic surface morphology, nine bionic concave surface micro-morphology have been designed on contact area of the ring model, three-dimensional finite element numerical simulation of nine bionic micro-morphology and smooth one was completed, comparative analysis of contact stress were completed by the stress nephogram of contact stress. It shows that the peak of smooth model is higher than bionic micro-morphology on contact stress on contact area, but the average contact stress of bionic concave surface micro-morphology ring model are lower smooth model.

Sixteen conventional evaluation-factors such as watershed area, maximum altitude difference wre chosen to evaluate the debris flow risk. Analyzing of the impact level of various factors on the debris flow and eliminated the correlation of the factors based on the minimum entropy analysis(MEA). According to the contribution rate to the debris system,the main factors and their weights were determined. On the basis of matter-element theory,extension set theory and dependent funetion ealeulation,a matter-element model evaluating the risk of debris flow.By practical calculation of dependent degree to the risk level of debris flow,the risk level is evaluated by extenics.Compared with the evaluation results of the literature,the debris flow evaluation results obtained by the extenics method can more precisely reflect the practical risk level of debris flow.This shows that the application of MEA to select the main factors and determine their weights has a certain rationality.

To evaluate the degradation property of the N,O-CMC/n-β-TCP composites with different mass proportion in simulated body fluid in vitro. The changes of pH, calcium concentration, phosphonium concentration,the weight, surface structure, components of the composites were detected at different time. The pH of SBF is over 6.9 in all composites, especially when the proportion is 1:1(pH:7.0-7.4). The weight increases with the degradation time.The bone-like apatite formed on the surface of all the composites and the largest number of bone-like apatite formed when the proportion is 1:1.

Sodium selenite was used as inorganic selenium source in Morchella mycelium enriched selenium submerged fermentation. The optimal fermentation parameters were researched. Selected culture temperature, fermentation time and selenium concentration as the optimized factors. Influence of levels of the factors on enriched selenium Morchella mycelium biomass and the selenium content were studied. Through the single factor tests and response surface method the optimal processing conditions of Morchella mycelium enriched selenium submerged fermentation was concluded. They are fermentation time 5.22d, culture temperature 25.73 ℃, selenium concentration 86.67 μg/ml; Prediction of mycelium biomass is 10.512 g/L, selenium content 18.8 μg/ml. Actual mycelium biomass is up to 10.339 g/L, selenium content 18.6 μg/ml. The results show that the response surface method is a most useful tool in Morchella selenium-rich submerged fermentation conditions optimization.

The surface tension of a series of dilution rate of chlorothalonil and mancozeb were measured. The feature of target plant leaves were observed. The pesticide wetting ability (PWAL)on leaves and the pesticide retentivity(PRL) of leaves were determined. The relationship between leaves feature and the behavior of pesticide were analysed.The results show that PRL and PWA are relevant on some degree. PRL is also affected by liquid adhesion on leaves . PWAL and PRL are affected by characteristics of the leaf surface.