| Person Re-identification is a retrieval task of matching the query person from multiple non-overlapping surveillance cameras,which has important applications in security,surveillance and investigation.Person re-identification generally includes two steps: feature extraction and feature matching.In general,person re-identification methods are based on images.However,there are some limits in person images,such as low resolution,different illumination,angle transformation and serious occlusion.These bring difficulty for feature extraction.Deep neural network is capable of extracting more discriminative and robust deep features from person images effectively.Meanwhile,it is found that person textual attribute can provide important semantic information based on person image,but few studies introduce person textual attribute into person re-identification.Therefore,combining person image and person textual attribute,this paper carries out research on person re-identification based on convolution neural network and graph convolution network.The main research contents and innovation of this paper are as follows.(1)This paper adopts traditional convolutional neural network in order to extract deep features from person image to focusing on the difficulty of feature extraction.Meanwhile,due to the closely connection between person textual attribute and person identity,which can provide important semantic information,this paper proposes a method named fusion with image and textual attribute based on convolution neural network for person re-identification.This method first extracts person image features and person textual attribute features,and finally fuses these features to obtain more discriminative features.(2)This paper also finds that there is an implicit tight correlation among person textual attributes,which can provide important semantic information for person identity.The graph convolution network is good at mining the potential semantic correlation information among nodes.Therefore,this paper proposes an improved method named fusion with image and textual attribute based on graph convolution network for person re-identification.This method regards person image and textual attribute features as graph nodes,and transmits graph node features through graph convolution operation,which makes the person image features fuse the implicit semantic correlation information among textual attribute,and finally obtains more robust person features.Comparing with the traditional methods,the experiment results show that the proposed method achieves better results on Market-1501 and Duke MTMC-re ID datasets. |
