Abstract: In the video surveillance, the limited available person images and unreliable data annotation led to the performance degradation of supervised person re-identification. To solve these problems, we proposed an unsupervised small sample person re-identification method that integrated Gabor features and convolution features. Gabor transform was used to extract multi-scale and multi-direction person texture and edge information, so as to realize the data augmentation of small sample person images in feature level. The redundant information was eliminated by feature encoding to improve the efficiency of feature similarity calculation. The convolutional auto-encoder network was adopted to extract the nonlinear deep convolution feature of pedestrian, which avoided the dependence of supervised learning algorithm on data annotation. The fusion of two heterogeneous features was applied to person similarity comparison, which implemented the feature augmentation of small samples and the improvement of person feature discrimination ability. Experiments were implemented based on Market-1501 and DukeMTMC-reID datasets, the rank-1 accuracy reached 74% and 67.1% respectively. The experimental results showed that the proposed network framework effectively improved the performance of small sample person re-identification.

Key words: Gabor transform, feature encoding, convolutional auto-encoder network, unsupervised learning, small sample person re-identification