To address the problem of low accuracy of oil and gas pipeline leakage identification, the digital twintechnology is introduced, and a digital twin pipeline leakage identification model is constructed based onarithmetic optimisation AOA-SVM(Arithmetic Optimization Algorithm-Support Vector Machine). Firstly, the 3DROM(3D Reduced Order Model) pipeline model of oil and gas pipelines is constructed using Ansys software.Secondly, the collected pipeline signals are imported into MySql database through Java interface, and then thedata are imported into the 3D ROM pipeline model. Finally, the AOA-SVM algorithm is used to carry out the work recognition of the pipeline signals in Matlab environment, and the recognition effect is shown in its dynamic form by Twin builder software. The recognition effect is shown in its dynamic form. In order to show the superiority of AOA-SVM condition recognition ability, it is compared with other popular SVM( Support Vector Machine) optimisation algorithms on the basis of the same signal. The comparison results show that AOA-SVM has the highest classification accuracy, which can reach 90. 5% , i. e. , the recognition model of the proposed digital twin can simulate the leakage of pipelines and has a high monitoring credibility.

In order to solve the problem that the Internet of Vehicles with limited computing power can not undertake a large number of real-time task computing, offloads vehicle tasks are introduced to the edge server for computing through MEC (Mobile Edge Computing), and a joint optimization scheme for the delay and energy consumption of vehicle task offloading is proposed based on the I-GWO( Improved Grey Wolf Optimizer). A computation offloading model constrained by computation delay, energy consumption, and edge server computing resources is established, and an offloading optimization problem with the goal of minimizing the total system consumption is proposed. By improving the GWO (Grey Wolf Optimizer ), the I-GWO used to solve optimization problem. Simulation results show that the proposed scheme can effectively reduce the total system consumption, and the convergence performance of I-GWO is greatly improved compared to GWO.

In response to the demand for higher precision and higher resolution magnetic measurement technology, especially for the improvement of the performance index of weak magnetic measurement technology, based on the principle of magnetic induction electromagnetic, using integrated fluxgate chip, a sensor component structure which can be used in weak magnetic three-dimensional vector field is designed. The influence of coil current, temperature and environment circuit of generating device of weak magnetic field is studied. The three- dimensional distribution characteristics of the magnetic field coil are analyzed and calculated by Ansys simulation software, and the magnetic field intensity and direction of the random space test point are determined. The heat dissipation of coil current is analyzed by COMSOL physical field calculation software. The experimental results show that the test data is basically consistent with the simulation results, so the excitation current condition of the weak magnetic field generator is determined. The system accurately measures the distribution and change of magnetic field vector of geomagnetic field with different latitudes and longitudes. The magnetic vector integral of 20 000 space points for the vector of any closed curve are realized, the curl theorem of static magnetic field is verified experimentally. 

In order to solve the energy limitation problem of oil and gas IoT, energy harvesting technology is introduced and an energy management algorithm for oil and gas IoT is proposed based on the data importance level. The working mode of sensor nodes is changed according to the energy threshold, and whether to transmit data is based on the importance level of data under the energy constrained situation to avoid the delay of important data. The algorithm selects cluster heads based on the ratio of energy harvesting and energy consumption of nodes, and the remaining energy of nodes, so as to achieve the purpose of extending the network lifetime while ensuring the transmission of important data. 

To address the issue of catastrophic forgetting in decision-making for autonomous driving in multi-city scenarios, a framework based on continual reinforcement learning is proposed. This framework is built upon the IMPALA( Importance Weighted Actor-Learner Architecture) algorithm architecture. First, a co-attentive awareness module is combined to extract critical environmental representations through cross-scenario feature interaction. Second, a self-activating neural ensemble architecture is built to enable autonomous activation of knowledge modules. Finally, a replay mechanism is applied to relieve the problem of forgetting old knowledge by combining scenario-specific features with historical trajectory experience replay. Off-policy behavior cloning and on-policy learning are employed concurrently to maintain the plasticity and stability of the decision-making algorithm. Whether to use old modules or generate new ones is determined based on the requirements of different autonomous driving scenarios and tasks, and the issue of excessive memory usage is addressed through module fusion. Ablation experiments and comparative ones are conducted in two different groups of multiple city scenarios. The performance of the proposed method is validated by comparing path completion rates and cumulative rewards. Experimental results demonstrate that the average completion rate reaches approximately 85% in the first sequential scenario, and it reaches 81. 93% in the second sequential scenario. The proposed scheme can effectively relieve the issue of catastrophic forgetting in multi-scenario continual decision-making and achieve better stable driving performance. 

Aiming at the problem that most of the existing pumping unit fault diagnosis is based on indicator diagram, which leads to a relatively single diagnostic modality, a ShuffleNetV2ECA-MLP (ShuffleNetV2 with Efficient Channel Attention and Multilayer Perceptron, ShuffleNetV2ECA-MLP) multimodal decision fusion fault diagnosis model is proposed for pumping units. In order to improve the cross-channel interaction capability and recognition accuracy of the ShuffleNetV2 model, firstly, the ECA(Efficient Channel Attention) module with lightweight channel attention is introduced into the ShuffleNetV2 model, and the Hardswish activation function is applied to enhance the network蒺s ability to learn complex problems. Secondly, the improved ShuffleNetV2 network is used to diagnose the figure of merit, and the MLP(Multi-Layer Perceptron) network is used to process the production dynamic data. Finally, the diagnostic results of the two models are integrated using the weighted voting method. In order to verify the effectiveness of the improved ShuffleNetV2 and ShuffleNetV2ECA-MLP models, comparisons are made with the lightweight convolutional networks MobileNetV2, MobileNetV3, the classical convolutional network ResNet, and the VGG ( Visual Geometry Group) network model. The experimental results show that the storage space of the ShuffleNetV2ECA-MLPmodel is only 10. 16 MByte, and the fault diagnosis accuracy reaches 95. 35% , which better meets the needs of pumping unit fault diagnosis.

Aiming at the problems of insufficient stability, weak generalization ability, and the influence of equipment configuration in the traditional scheduling method for the detection and scheduling of power metering equipment, a detection and scheduling model named GNN-RL(Graph Neural Network-Reinforcement Learning) is proposed. The model treats the scheduling problem as a Markov decision process. Firstly, the graph structure model of electric energy metering equipment detection and scheduling is constructed. Then, the problem features are extracted through the improved graph neural network and passed to the action selection network to generate decisions. After the scheduling, the model collects feedback information to train the scheduling policy in the reinforcement learning module. In the training phase, GNN-RL optimizes the message passing mechanism, employs a loss function closely related to the scheduling objective, and dynamically adjusts the learning rate. A multi-task learning framework is introduced to deal with task allocation and time scheduling. The experimental results show that GNN-RL has obvious advantages in optimization ability, solution accuracy, and stability, and has great advantages in solving the detection and scheduling problem of energy metering equipment, which significantly improves the efficiency and reliability of problem solving.

Aiming at the nonlinear problem of SMC (Sliding Mode Control) of PMSM (Permanent Magnet Synchronous Motor), on the basis of analyzing the weak connection between the parameters of the sliding mode control at different stages, a fuzzy parameter tuning strategy is proposed to reduce overshoot, chattering and having certain anti-jamming ability. The major functions of various parameters and the relationship between parameters in the process of sliding mode motor control are improved. The fuzzy logic control is combined with the exponential reaching law. The change of sliding mode parameters is adjusted by fuzzy control, and the fuzzy reaching law is derived. The design method simplifies the complexity of sliding mode parameter adjustment and eliminates the constant speed term of exponential approach rate, making the design and regulation more convenient. The conclusions show that compared with the traditional sliding mode control of permanent magnet synchronous motor, the speed loop controller designed by making fuzzy control to adjust the sliding mode parameters can further reduce the overshoot by 21. 86%, reduce the start-up time by 0.038 s, effectively eliminate the buffeting peak wave when reducing the buffeting, and reduce the speed drop by 7. 31% when dealing with the sudden change of load. The adjustment time is reduced by 0. 010 s. The reliability and superiority of the parameters are verified. 

SLAM(Simultaneous Localization and Mapping) technology is a camera pose estimation based on static scene features. It is susceptible to dynamic objects in the process of feature calculation and matching at its front end. Therefore a method of instance segmentation combined with multi-view geometric constraints is proposed to improve the front-end feature processing of visual SLAM and eliminate the interference of dynamic information. Specifically, in the front end of the ORB-SLAM3(Oriented FAST and Rotated BRIEF-Simultaneous Localization and Mapping3) framework, the YOLACT + + (You Only Look At CoefficienTs + +) instance segmentation thread is paralleled, and the segmented results are used to supplement the multi-view geometric constraint method testing the dynamic consistency of feature points. The EfficientNetV2 network is used to replace the original backbone network of YOLACT + +, and the TensorRT is used to quantify the instance segmentation model to reduce the front-end computing pressure of the algorithm. The test of TUM data set shows that the positioning accuracy of the proposed algorithm in high dynamic environment is 80. 6% higher than that of ORB-SLAM3 algorithm. 

To improve the efficiency and accuracy of the helmet wearing detection model, the LFE-Y8 (LightConv, Focal Loss and EMA Attention You Only Look Once version 8) model is proposed. This model adopts the Focal Loss function to solve the problem of imbalanced sample categories. The original model is optimized using LightConv lightweight convolution, which improves the feature extraction ability. In order to better focus on small targets, an efficient multi-scale EMA (Efficient Multi Scale Attention) attention mechanism for cross spatial learning is integrated. The experimental results show that the LFE-Y8 model effectively improves the accuracy of helmet wearing detection compared to the improved YOLOv8 model. The improved algorithm has an accuracy increase of 0. 6% and a recall increase of 2. 1%. The mAP@ 50 is improved by 1.2%, and mAP@ 50-95 is improved by 1.5%, demonstrating the effectiveness of the LFE-Y8 model in practical applications. 

Aiming at the high energy consumption problem that exists in multiple-input multiple-output C-IoT (Cognitive Internet of Things) systems, a joint beamforming optimization algorithm based on IRS(Intelligent Reflecting Surface) assistance is proposed. Taking the signal-to-interference-to-noise ratio at the secondary user and the discrete phase shift at the IRS as constraints, a new optimization criterion is constructed to maximize the energy efficiency of system by jointly optimizing the active beamforming matrix at the secondary transmitter and the passive beamforming matrix at the IRS. After decomposing the complex non-convex optimization problem into sub-problems, the fixed-point iteration method and the successive refinement method are used to process the sub- problems respectively. The simulation results show that the proposed algorithm has good convergence in multi- antenna scenarios. Compared to the baseline scheme, the proposed algorithm effectively improves the energy efficiency of the system in multi-user situations.

With the continuous integration of large-scale distributed power sources into distribution networks, distribution networks face many challenges in terms of safety, stability and economy. And the existing deep reinforcement learning methods often exhibit limitations in generalization ability when training agents to cope with changing operating conditions due to insufficient generalization of collected data. Therefore a distribution network regional voltage optimization control strategy based on large language model-assisted deep reinforcement learning is proposed, combining large language model technology with deep reinforcement learning. Secondly, by guiding large language models to generate customized datasets for deep reinforcement learning agent training through prompt engineering, a multi-agent collaborative decision-making framework is constructed. Then, based on distributed partially observable Markov processes, dynamic control problems are modeled to reduce dependence on real-world data while improving agent generalization ability. Finally, the effectiveness of the proposed control strategy is verified on the improved IEEE 33-node system, with voltage deviation and network loss reduced by 60. 82% and 49.91%, respectively, exhibiting strong robustness under various operating conditions.

In order to study the conjecture of neighbor sum distinguishing total chromatic number, the structure of minimal counter example graphs is analyzed using the Combinatorial Nullstellensatz and proved that the neighbor sum distinguishing total chromatic number of IC-planar graphs with the maximum degree Δ>=8 without intersecting triangles does not exceed Δ+3 by the discharging method. The research results indicate that the conjecture of neighbor sum distinguishing total chromatic number holds on this class graphs.

The accuracy of the existing estimation methods of human pose in the motion evaluation scene needs to be further improved. The methods rely on high-performance computing devices, and the reasoning speed on edge computing devices needs to be further enhanced. Therefore, improvement is made to the classic high-resolution network model to solve the problem of low real-time performance of the existing human pose estimation methods. To address the frequent occlusion issues in motion evaluation scene, random erasure enhancement is applied to the images in the dataset. After experimental comparison and verification, the improved method significantly reduces the number of model parameters and improves the inference speed of the model while ensuring the accuracy of attitude estimation. The algorithm exhibits stronger robustness for occlusion problems, and the improved method can meet the needs of motion evaluation scenarios.

In the coal gangue sorting scene, coal and gangue are closely intertwined, making it challenging for the mechanical claw to accurately grasp gangue. To enhance grasping reliability under complex conditions, a multi-objective small-gap coal gangue recognition method based on semantic segmentation is proposed. The FCN (Fully Convolutional Networks) algorithm is improved by integrating the FPN (Feature Pyramid Network) module and replacing the cross-entropy loss with Dice Loss. A total of 1 202 coal gangue images are annotated using Labelme, and enhanced with the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm. The performance of various semantic segmentation algorithms and transfer learning strategies are analyzed by comparative experiment. Results show that the FCN-ResNet50-FPN model, with ResNet50 as the feature extractor, achieved a precision of 95. 0%, a recall of 95. 4%, an F1 score of 95. 2%, and a mIoU(mean Intersection over Union) of 90. 9%. Transfer learning further improves recognition, enhances small-gap detection and provides reliable data for precise coal gangue sorting robot operations. 

Images in complex backgrounds may contain a large amount of interference information, making it difficult to accurately extract image feature vectors. Therefore, a salient target recognition algorithm for wireless sensor network images in complex backgrounds is proposed. Mean shift algorithm is used to cluster targets and achieve image segmentation, obtaining several segmented image feature vectors. By fusing rules, the pyramid shapes of each layer in the image pyramid are prosessed, and the image pyramid sequence is combined to achieve image reconstruction. The image saliency of pixels is determined based on the results of image reconstruction and histogram statistics, and image saliency target recognition is achieved through pixel order expansion and sorting. The experimental results show that the proposed algorithm can accurately identify salient targets in wireless sensor network images under complex backgrounds, and the texture details of the image are clear, with good practical application effects. 

Data may be tampered with by illegal elements during storage, leading to security issues such as data leakage and damage. In order to ensure the security of data and further improve the anti tampering ability of cloud data, an anti tampering storage algorithm for multi blockchain cloud data is proposed. By analyzing the structure of multi blockchain networks, a cloud data distributed storage architecture based on blockchain technology is designed. A cloud data access control model is designed through the data layer, application interaction layer, contract layer, and functional layer to improve the security of distributed cloud data storage. During the data encryption and decryption process, Hash algorithm and asymmetric encryption algorithm are combined to encrypt and decrypt the combined data of “random number+cloud data冶, to avoid the phenomenon of node private key leakage, cloud data being replaced and forged. The experimental results show that the encryption and decryption processing time of the proposed algorithm is maintained within 25 ms, the probability of successful attack is less than 40%, and the data is not easily tampered with during transmission.

In response to the demand for real-time and accurate tracking of multiple surgical instruments in collaborative operations, this study aims to solve the problem of invisible instrument ends caused by light filtering and shielding based on a near-infrared active optical tracking system, and achieve accurate identification and stable tracking of multiple surgical instruments. The problem of surgical instrument tracking based on a near- infrared optical positioning system is explored. In response to the limitations of traditional algorithms in multi- instrument recognition and landmark tracking, a multi-instrument recognition method based on point set matching is proposed. By matching the landmark points in the captured image with the pre-established instrument model, accurate identification and differentiation of multiple surgical instruments are achieved. And a landmark tracking method based on least squares algorithm prediction is proposed. The historical tracking data and the least squares method are used to predict the location of the landmark points, thereby achieving real-time tracking of surgical instruments. Experimental verification shows that the coordinated use of the two methods can achieve accurate identification (first-frame recognition rate 100% ) and stable tracking (overall recognition rate 97. 45%) of multiple surgical instruments, while ensuring tracking accuracy and meeting the real-time requirements of surgery. It can provide surgeons with clear instrument spatial position information and support precise operations. It is of great significance to improving surgical safety and reliability, and can provide technical support for the clinical application of near-infrared optical surgical navigation systems.

Due to the complexity of multi tenant and multi-level security requirements in cloud computing environments, existing access control strategies are difficult to meet the needs of different users and applications, resulting in lower resource access security and more time consumption during encryption and decryption. To address the above issues, combining CP-ABE(Ciphertext Policy-Attribute Based Encryption) with asymmetric encryption RSA(Rivest Shamir Adleman) research on hierarchical access control of cloud computing resources is conducted. A CP-ABE access control architecture is established and encrypted access policies are developed. The trust values related to hierarchical access of cloud computing resources is used to clarify the trust relationship of access, and the trust degree of access subject and object resources are obtainined. Based on the calculated trust value, users are granted hierarchical authorization to meet the needs of multi tenant and multi-level security. Based on the results of user identity hierarchical authorization, the RSA algorithm is used to replace the complex bilinear mapping of CP-ABE for encryption and decryption, achieving precise control of resourcehierarchical access and reducing encryption and decryption time consumption. Through experimental testing, it was found that the proposed method can achieve a concurrent connection count of 400, a maximum over authorization rate of 6. 8% for hierarchical access, and an effective control of access control response time of less than 6 seconds, which can effectively meet the multi tenant and multi-level security needs of application scenarios. It has a good effect on hierarchical access control of cloud computing resources. 

Due to the poor defogging effect of remote sensing image clustering, the clustering accuracy and Kappa coefficient of image clustering are low and the time is long. In order to solve these problems, a new clustering optimization method remote sensing image based on adaptive K-means algorithm is proposed. Firstly, dark channel prior estimation and color line prior estimation are used to de-fog remote sensing images. Secondly, the gray co-occurrence matrix of the remote sensing image after fog removal is calculated, and the texture features are obtained. Finally, the colony algorithm is used to optimize the K-means algorithm, and the optimized adaptive K-means algorithm is used to realize the clustering optimization of remote sensing images according to texture features. The experimental results show that the proposed method can effectively eliminate cloud and fog in remote sensing images, and the image details are clearly displayed. The proposed method has good performance in terms of clustering accuracy, Kappa coefficient and clustering time. The clustering accuracy reaches 94. 9%, the Kappa coefficient is 0. 97, and the clustering time is 0.36 s. This method has certain validity. 

 In order to quickly extract key information from massive library knowledge graphs, an entity relationship extraction algorithm for academic text keyword library knowledge graphs is proposed. OCS-FCM (Optimization of Complete Strategy Fuzzy C-Means) and Elastic E-t-SNE(Embedding t-Distributed Stochastic Neighbor Embedding ) algorithms are used to perform missing value filling and dimensionality reduction on key words in the library. And using entities in the academic text keyword library as vertices, a knowledge graph is established. Based on the part of speech and other features of keywords, a SelfATT BLSTM(Self Attention Bidirectional Long Short Term Memory) model is constructed using a self attention mechanism algorithm to extract entity relationships from the knowledge graph and obtain the extracted results. Experimental results have shown that the collection accuracy of proposed algorithm is more than 0. 8, with an ACC(Accuracy) value over 30% and a extraction time less than 1.5 s, demonstrating excellent ability to extract entity relationships. 

In order to improve the accuracy of video anomaly detection in complex scenes, a video anomaly detection framework based on improved GAN(Generative Adversarial Network) is proposed. Two discriminators are used for the adversarial training of the generator, and the bidirectional prediction consistency is enhanced through a regression loss function. FusionNet and LSTM(Long Short Term Memory) are combined to form a generator structure based on spatio-temporal feature fusion. Forward and backward video sequences are taken as the inputs of the generator, and predicted video frames and predicted video sequences are output respectively. Patch GAN architecture is adopted for both of the discriminators, the frame discriminator is used to distinguish synthetic frames and the sequence discriminator is used to determine whether the frame sequence contains at least one synthetic frame to maintain temporal consistency of the predicted frames, to improve the robustness and accuracy of the predicted network. Finally, the anomaly score is calculated based on the normalized mean PNSR (Peak Signal to Noise Ratio). Experimental results show that the proposed framework can effectively capture the bidirectional spatio-temporal features in video sequences and outperforms other state-of-the-art methods on thechallenging public video anomaly detection datasets UCF-Crime ( University of Central Florida Crime) and ShanghaiTech.

Text classification of transmedia knowledge involves many types of data, such as text, image, video, etc. The heterogeneity and heterogeneity of data increase the complexity of classification. Aiming at the problem that it is difficult to find accurate data in a large number of cross-media knowledge texts, an algorithm of cross- media knowledge text classification based on multi-modal information is proposed. The TF-IDF(Term Frequency- Inverse Document Frequency) algorithm is used to filter the stop words in the processing text, extract the text features, and integrate them with the image text features. By using naive Bayes classifier, the classification of cross-media knowledge text is determined and realized. Experimental analysis shows that the proposed text classification algorithm significantly improves the performance and efficiency of cross-media knowledge text classification, and makes the classification results more accurate, with the accuracy rate up to 95. 12% and the missing rate remaining below 10%.

The sources of regional economic data are diverse, including statistical departments, enterprise reports, sensor data, et al. There are significant differences in data format, structure, and semantics, making it difficult to process them uniformly. This leads to difficulties in accurately extracting data features, which in turn results in inaccurate data classification results for methods. To address this issue, a regional economic big data integrated classification method based on parallel clustering algorithm is proposed. Based on the characteristics of regional economic big data, calculate the purity and neighborhood radius of the data, determine the missing values of regional economic big data, and correct and fill them in. Based on the filled data, parallel clustering algorithm is used to randomly divide it into multiple subsets of data. The parallel clustering algorithm utilizes multi node parallel processing to significantly improve computational efficiency and meet the requirements of large-scale data processing. Extract the feature quantities of each data subset and design a big data base classifier accordingly. Under the premise of considering the internal data density of the base classifiers, determine the weight values of each base classifier, combine the classification results of each base classifier, and output the final data ensemble classification result. The experimental results show that the designed classification method has a DBI (Davies-Bouldin Index) index of 0.31 in practical applications, which can achieve accurate classification of regional economic big data.

Multi source heterogeneous data may contain sensitive and personal privacy information, increasing the risk of data leakage. Therefore, a multi-source heterogeneous data security aggregation algorithm based on kernel extreme learning machine is designed. Partial least squares algorithm is used to extract features from multi-source heterogeneous data, the extreme learning machine is optimized by introducing kernel functions, and the obtained data features inputtied into the kernel extreme learning machine to complete data aggregation by class. Elliptic curve encryption algorithm is used to encrypt the aggregated data, improving data security and achieving the goal of secure aggregation of multi-source heterogeneous data. The experimental results show that the algorithm has high accuracy in multi-source heterogeneous data aggregation and good data encryption performance, and can be widely applied in practice. 

To ensure the confidentiality and integrity of sensitive data for smart grid end users, a secure aggregation algorithm for smart grid end user sensitive data is proposed. The establishment of a fog computing architecture is prioritized for sensitive data collection, transmission, and processing of smart grid end users. At the smart meter terminal, the collected user sensitive data is encrypted through the formation and distribution of keys. The encrypted data is aggregated through fine-grained aggregation of fog nodes and coarse-grained aggregation of cloud nodes, and the obtained coarse-grained aggregation dataset of cloud nodes is transmitted to the power service institution through secure transmission channels. By parsing all fine-grained aggregated data stored in the cloud, the corresponding aggregated data plaintext can be obtained to achieve secure aggregation of sensitive data for smart grid end users. The experiment shows that the communication overhead of the proposed algorithm for secure aggregation of sensitive data is below 100 kByte, and the aggregation time is short. And the integrity protection for different types of sensitive data is above 0. 8, indicating that the proposed algorithm has high practicality.

A real-time image feature tracking algorithm based on visual communication effect is proposed to address the issues of low extraction accuracy and poor tracking performance in real-time tracking of multi frame film and television images. By calculating the likelihood function of candidate target states, effective targets in multiple film and television images are obtained, and unnecessary information such as noise and artifacts in the images are removed through visual communication enhancement processing to improve the recognizability of effective targets. Considering the real-time tracking requirements of multi frame film and television images, after multi frame film and television image enhancement, particle swarm optimization algorithm is used to weight the features of multi frame film and television images, improve the weight of salient features, effectively highlight the tracking target, and achieve real-time tracking of multi frame film and television image features. The experimental results show that after processing with the proposed method, the signal-to-noise ratio of the image can reach 45 dB, and the position of the tracked target is relatively consistent with the original marked position. And the target recognition rate, recall rate, and comprehensive evaluation index are all maintained above 0. 9. It is demoustrated that the proposed method can filter out the interference of target image background and surrounding clutter, accurately obtain target feature points, and achieve real-time feature tracking of multiple film and television images. 

In low light images, the effective signal and random noise exhibit a similar sparse distribution in the transform domain. The denoised image is prone to staircase effects or pseudo edges, which can lead to noise artifacts and reduce the quality of the image. Therefore, an online dictionary learning algorithm is proposed to smooth and denoise low light images, and to improve the visual effect of the images. The grayscale transformation on low light images is Implemented to reduce random noise in the images. An adaptive low light image block partitioning strategy is designed which dynamically adjusts the size of image blocks based on local brightness information and texture features after grayscale transformation, to obtain image details and structural information. An online dictionary learning model is created. It sparsely represents partitioned image blocks, dynamically captures the time-varying characteristics of noise and detail features through real-time dictionary updates, adaptively separates effective signals and noise, suppresses noise artifacts while preserving signal structure, solves the problems of staircase effect and pseudo edges caused by similar sparse distribution in low light images, and achieves smooth denoising of low light images. The experimental results show that the proposed method has strong robustness and can effectively suppress low light image noise. The peak signal-to-noise ratio and structural similarity of the image are significantly improved. 

Considering the large-scale and high-dimensional features of remote sensing images, the complex remote sensing application process requires appropriate feature extraction and selection, while further distinguishing different subcategories of similar targets. Therefore, an improved ResNet(Residual Network) algorithm is proposed for fine-grained recognition of complex remote sensing background targets. Non mean filtering algorithm is used to label the coordinate domain of noisy remote sensing images, calculate the similarity between pixels, and denoise complex remote sensing images. Based on the denoising results, global and local feature points of the image are extracted, and global and local feature maps are obtained through feature point fusion results. An improved residual network algorithm is introduced to analyze the fine-grained pixel size of each background image block area. After residual learning, combined with the image pixel position and loss function, a classifier is used twice to determine the fine-grained pixel features and complete the fine-grained recognition of background targets. The experimental results show that the image clarity is high, and as the number of images to be recognized continues to increase, F1 -Score and global recall rates have been improved to various degrees, with lower gain errors. 

To enhance the security and confidentiality of patients’ medical privacy information, a secure encryption method based on chaotic mapping and the RC5(Rivest Cipher 5) algorithm is proposed. Firstly, the PCA(Principal Component Analysis) method is used to perform data dimension specification processing on medical privacy information data, completing dimensionality reduction of medical privacy information data. Then, using the RC5 block cipher algorithm, the key is initialized, transformed, and mixed to achieve secure encryption of medical privacy information. Finally, the dual chaotic system encryption algorithm ( EDC: Enhanced Dual Chaos) is used, combined with the high-quality performance of one-dimensional logistic mapping and two-dimensional Henon mapping, to achieve secondary encryption of medical privacy information. The experimental results show that the proposed method has high encryption efficiency, low information leakage rate, and strong anti attack performance.

A new method for accurately identifying key information in financial bill images under the interference of stains is proposed to solve the problem of recognition failure caused by local information loss or misjudgment during the recognition of financial bill information containing stains. By integrating the gray scale and denoising of the image, dividing the bill area according to the threshold size, binary processing is achieved. Using the ROI (Region of Interest) principle, rough screening of information range is performed in the binary processed image to calculate the maximum coverage contour and key information adhesion area. Further segmentation and edge calibration are performed, and key information areas are located based on the ROI ratio. The features of key information areas is extracted, classification labels in an end-to-end manner are established by combining the features, and the recognition target is searched according to the label gradient in the coverage recognition area. The experimental results show that the retrieval rate obtained by the proposed method is between 80%-100%, and the efficiency is greatly improved with superior and reliable performance.