Semi-supervised dictionary learning based deep network for person search

Semi-supervised dictionary learning based deep network for person search

연구 분야: Artificial Intelligence

학회: Multimedia Systems

Current person search methods are divided into two categories, i.e., supervised and weakly supervised methods. The supervised methods suffer from the burden of manual annotations, and the weakly supervised methods sacrifice the accuracy. How to reduce manual annotations and improve performance is still a challenge for person search. In the paper, we propose an end-to-end semi-supervised dictionary learning based deep network (SSDLNet) for person search. In SSDLNet, we improve the SSD512 model for person detection sub-task and feature extraction to solve the problem of insufficient deep information learning and expression in the original SSD512, and design a person re-identification (re-ID) module to receive the person features output from the improved SSD512 model and generate the re-ID results. In the person re-ID module, we present semi-supervised dictionary learning (SSDL), where we construct a structured dictionary that contains several sub-dictionaries, each corresponding to a person identity class. SSDL calculates the sparse representations of person features to re-identify the individuals. SSDLNet makes full use of both labeled and unlabeled person images for training. Experimental results show that SSDLNet achieves good performance on CUHK-SYSU and PRW datasets, as compared with state-of-the-art methods.