基于滑动特征向量的小样本图像分类方法

doi:10.13229/j.cnki.jdxbgxb20200532

Abstract

Abstract:

In the task of few-shot image classification， the extremely limited number of labeled examples per class can hardly represent the real class distribution effectively， which is the main reason for misclassification. To tackle this problem， we propose a method which named Sliding Feature Vectors Neural Network （SFV）. The method aims to assemble all the local sliding feature vectors of samples from the same class to construct the class-level feature spaces， and then it utilized the image-to-class measure to classify the query samples. That means on the measure stage， SFV compare the similarity between the class and the query sample. SFV expands the class feature space by adding the edge information of feature blocks and correlation of their position and structures to maximize the utilization of the deep feature maps when the sample is extremely limited， which can ease overfitting problem caused by small sample data. Experimental study on benchmark datasets consistently shows its superiority over the related other framework， especially on fine-grained datasets， it achieves state-of-the-art.

Key words: computer application technology, computer vision, few-shot learning, local features, metric learning

CLC Number:

TP183

Jie CAO,Xue QU,Xiao-xu LI. Few⁃shot image classification method based on sliding feature vectors[J].Journal of Jilin University(Engineering and Technology Edition), 2021, 51(5): 1785-1791.

Figures/Tables 6

Fig.1

Table 1

Table 2

Table 3

Table 4

Table 5

References 18

1	车翔玖, 董有政. 基于多尺度信息融合的图像识别改进算法[J]. 吉林大学学报: 工学版, 2020, 50(5): 1747-1754.
	Che Xiang-jiu, Dong You-zheng. Improved image recognition algorithm based on multi⁃scale information fusion[J]. Journal of Jilin University(Engineering and Technology Edition), 2020, 50(5): 1747-1754.
2	Wang Y, Yao Q, Kwok J T, et al. Generalizing from a few examples: a survey on few-shot learning[J].ACM Computing Surveys (CSUR), 2020, 53(3): 1-34.
3	Finn C, Abbeel P, Levine S. Model-agnostic meta-learning for fast adaptation of deep networks[C]∥International Conference on Machine Learning, Sydney, 2017: 1126-1135.
4	Ravi S, Larochelle H. Optimization as a model for few-shot learning[C]∥International Conference on Learning Representations, San Juan, 2016: 1-11.
5	Chen W Y, Liu Y C, Kira Z, et al. A closer look at few-shot classification[C]∥International Conference on Learning Representations, New Orleans, 2019: 04232.
6	Cai Q, Pan Y, Yao T, et al. Memory matching networks for one-shot image recognition[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, 2018: 4080-4088.
7	Koch G, Zemel R, Salakhutdinov R. Siamese neural networks for one-shot image recognition[C]∥International Conference on Machine Learning, Lille, France, 2015.
8	Vinyals O, Blundell C, Lillicrap T, et al. Matching networks for one shot learning[C]∥Neural Information Processing Systems, Barcelona, 2016: 3630-3638.
9	Snell J, Swersky K, Zemel R. Prototypical networks for few-shot learning[C]∥Neural Information Processing Systems, Long Beach, 2017: 4077-4087.
10	Sung F, Yang Y, Zhang L, et al. Learning to compare: relation network for few-shot learning[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Salt Lake City, 2018: 1199-1208.
11	刘萍萍, 赵宏伟, 耿庆田, 等. 基于局部特征和视皮层识别机制的图像分类[J]. 吉林大学学报: 工学版, 2011, 41(5): 1401-1406.
	Liu Ping-ping, Zhao Hong-wei, Geng Qing-tian, et al. Image classification method based on local feature and visual cortex recognition mechanism[J]. Journal of Jilin University(Engineering and Technology Edition), 2011, 41(5): 1401-1406.
12	Li W, Wang L, Xu J, et al. Revisiting local descriptor based image-to-class measure for few-shot learning [C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, 2019: 7260-7268.
13	Russakovsky O, Deng J, Su H, et al. Imagenet large scale visual recognition challenge[J]. International Journal of Computer Vision, 2015, 115(3): 211-252.
14	Welinder P, Branson S, Mita T, et al. Caltech-UCSD birds 200[R]. Technical Report CNS-TR-2010-001, California Institute of Technology, 2010: 1-15.
15	Krause J, Stark M, Deng J, et al. 3D object representations for fine-grained categorization[C]∥Proceedings of the IEEE International Conference on Computer Vsion Workshops, Sydney, 2013: 554-561.
16	Khosla A, Jayadevaprakash N, Yao B, et al. Novel dataset for fine-grained image categorization: stanford dogs[C]∥Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Colorado Springs, USA, 2012: 3181866.
17	Oreshkin B N, Rodriguez P, Lacoste A. TADAM: task dependent adaptive metric for improved few-shot learning[C]∥Neural Information Processing Systems, Canada, 2018: 721-731.
18	Rusu A A, Rao D, Sygnowski J, et al. Meta-learning with latent embedding optimization[C]∥International Conference on Learning Representations, New Orleans, 2019: 05960.

Related Articles 15

[1]	Chun-bo WANG,Xiao-qiang DI. Cloud storage integrity verification audit scheme based on label classification [J]. Journal of Jilin University(Engineering and Technology Edition), 2021, 51(4): 1364-1369.
[2]	Rong QIAN,Ru ZHANG,Ke-jun ZHANG,Xin JIN,Shi-liang GE,Sheng JIANG. Capsule graph neural network based on global and local features fusion [J]. Journal of Jilin University(Engineering and Technology Edition), 2021, 51(3): 1048-1054.
[3]	Qian-yi XU,Gui-he QIN,Ming-hui SUN,Cheng-xun MENG. Classification of drivers' head status based on improved ResNeSt [J]. Journal of Jilin University(Engineering and Technology Edition), 2021, 51(2): 704-711.
[4]	Yuan SONG,Dan-yuan ZHOU,Wen-chang SHI. Method to enhance security function of OpenStack Swift cloud storage system [J]. Journal of Jilin University(Engineering and Technology Edition), 2021, 51(1): 314-322.
[5]	Tao XU,Ke MA,Cai-hua LIU. Multi object pedestrian tracking based on deep learning [J]. Journal of Jilin University(Engineering and Technology Edition), 2021, 51(1): 27-38.
[6]	Xiang-jiu CHE,You-zheng DONG. Improved image recognition algorithm based on multi⁃scale information fusion [J]. Journal of Jilin University(Engineering and Technology Edition), 2020, 50(5): 1747-1754.
[7]	Hong-wei ZHAO,Ming-zhao LI,Jing LIU,Huang-shui HU,Dan WANG,Xue-bai ZANG. Scene classification based on degree of naturalness and visual feature channels [J]. Journal of Jilin University(Engineering and Technology Edition), 2019, 49(5): 1668-1675.
[8]	CHE Xiang-jiu, WANG Li, GUO Xiao-xin. Improved boundary detection based on multi-scale cues fusion [J]. Journal of Jilin University(Engineering and Technology Edition), 2018, 48(5): 1621-1628.
[9]	XU Yan-yan, CHEN Hui, LIU Jia-ju, YUAN Jin-zhao. Cell processor stereo matching parallel computation [J]. 吉林大学学报(工学版), 2017, 47(3): 952-958.
[10]	HU Guan-yu, QIAO Pei-li. High dimensional differential evolutionary algorithm based on cloud population for network security prediction [J]. 吉林大学学报(工学版), 2016, 46(2): 568-577.
[11]	ZHANG Bao-hua, HUANG Wen-qian, LI Jiang-bo, ZHAO Chun-jiang, LIU Cheng-liang, HUANG Dan-feng. Online sorting of irregular potatoes based on I-RELIEF and SVM method [J]. 吉林大学学报(工学版), 2014, 44(6): 1811-1817.
[12]	YANG Yan, LIU Sa, LIAN Shi-bin, ZHU Xiao-dong. Analysis of fruit tree pests morphological characteristics based on computer vision [J]. 吉林大学学报(工学版), 2013, 43(增刊1): 235-238.
[13]	TONG Jin, WANG Ya-hui, FAN Xue-mei, ZHANG Shu-jun, CHEN Dong-hui. Monitoring system of cold chain logistics for farm fresh produce [J]. 吉林大学学报(工学版), 2013, 43(06): 1707-1711.
[14]	SHANG Fei, MA Jun-Xiao, YAO Li, TIAN Di, QIU Chun-Ling. Multifeature fusion based method for monitoring working status of instruments [J]. 吉林大学学报(工学版), 2010, 40(02): 545-0548.
[15]	GE Liang, ZHU Qing-sheng, FU Si-si, LUO Da-jiang, LIU Jin-feng. Improved image dense stereo matching algorithm [J]. 吉林大学学报(工学版), 2010, 40(01): 212-0217.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

Comments

Recommended 0

No Suggested Reading articles found!

模型		嵌入模块	5-way 1-shot	5-way 5-shot
基线实验	DN4	Conv-64F	51.24±0.74	71.02±0.64
	DN4*	Conv-64F	51.78±0.80	70.15±0.73
	全局特征KNN	Conv-64F	44.54±0.76	49.52±0.75
基于度量学习方法	匹配网络*	Conv-64F	43.56±0.84	55.31±0.73
	原型网络*	Conv-64F	48.45±0.96	66.53±0.51
	关系网络*	Conv-64F	50.44±0.82	65.32±0.70
	SFV	Conv-64F	53.81±0.81	71.98±0.68
基于元学习方法	Baseline*	Conv-64F	36.34±0.58	54.88±0.67
	Meta-Learner LSTM	Conv-32	43.44±0.77	60.60±0.71
	MAML*	Conv-32F	48.70±1.84	63.11±0.92
	TADAM	ResNet-12	58.50±0.30	76.70±0.30
	MM-Net	Conv-32F	53.37±0.48	66.97±0.35
	LEO	WRN-28-10	61.76±0.08	77.59±0.12

模型		Stanford Cars		Stanford Dogs		CUB-200
模型		5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot	5-way 1-shot	5-way 5-shot
基于度量学习方法	全局KNN	39.08±0.77	41.61±0.71	41.92±0.82	45.91±0.82	41.53±0.81	49.01±0.80
	DN4	59.84±0.80	88.65±0.44	45.41±0.76	63.51±0.62	46.84±0.81	74.92±0.64
	DN4*	60.73±0.85	88.45±0.49	48.90±0.82	69.63±0.74	54.28±0.92	74.66±0.75
	匹配网络*	34.80±0.98	44.70±1.03	35.80±0.99	47.50±1.03	45.30±1.03	59.50±1.01
	原型网络*	40.90±1.01	52.93±1.03	37.59±1.00	48.19±1.03	37.36±1.00	45.28±1.03
	SFV	65.18±0.85	88.39±0.48	52.19±0.86	69.86±0.72	57.81±0.93	75.59±0.74
基于元学习方法	MAML*	44.34±0.81	61.42±0.71	43.64±0.85	56.62±0.75	46.55±0.88	63.20±0.79
基于元学习方法	Baseline*	28.53±0.52	40.41±0.58	29.58±0.49	42.24±0.63	32.00±0.58	51.01±0.69

模型	CUB-200
模型	5-way 1-shot	5-way 5-shot
全局KNN网络（I）	41.53±0.99	49.01±0.80
全局KNN-全局平均池化网络（II）	41.04±0.99	47.35±0.92
特征块网络（III）	54.55±0.94	69.75±0.80
滑动特征块网络（IV）	56.25±0.92	74.37±0.76
滑动特征向量网络SFV（V）	57.81±0.93	75.59±0.74

a的取值	5-way 1-shot	5-way 5-shot
a=1	54.28±0.92	74.66±0.75
a=3	57.81±0.93	75.59±0.74
a=5	57.58±0.95	73.46±0.75

模型	5-way 1-shot	5-way 3-shot	5-way 5-shot
全局KNN*	41.53±0.81	48.71±0.91	49.01±0.80
DN4^*	54.28±0.92	68.72±0.82	74.66±0.75
SFV	57.81±0.93	70.90±0.83	75.59±0.74

Few⁃shot image classification method based on sliding feature vectors

RICH HTML

PDF (PC)