| The vehicle re-identification(vehicle re-ID)refers to selecting all images of the same vehicle captured by different cameras with non-overlapping views from a large database.Existing vehicle re-ID methods extract features from the visual appearance of vehicles,but these methods do not ignore the background and other irrelevant information in images during feature extraction,which makes it difficult for the model to obtain better recognition accuracy.In addition,vehicle images in existing datasets can’t contain all viewpoints,and these methods ignore the relationship between the vehicle images,resulting in insufficient representation of vehicle features extracted from a single viewpoint.Aiming at the above problems,this paper proposes the global and spatial multi-scale contexts fusion for vehicle re-identification,and the vehicle re-identification algorithm based on relation mining among images.The research contents and results of this paper are as follows:(1)Global and spatial multi-scale contexts fusion for vehicle re-identification.Due to the existence of background redundancy in images taken in complex scenes,it is difficult for the model to obtain strong discriminant features.In this paper,we propose a novel method called global and spatial multi-scale contexts fusion for vehicle re-identification.Firstly,a global contextual selection modul is designed,which divides the original feature map into several parts along the spatial dimension and learns the importance score of each part,so as to enhance the response of key regions and extract discriminative detailed features.Secondly,a multi-scale spatial context feature selection module is designed,which adopts multi-scale division for the optimized features and selectively optimize those multi-scale features to generate the foreground probability map of the vehicle image.At the same time,this module can remove the influence of noise and redundant information,improve the perception ability of vehicle spatial position features.Finally,through the cooperation of the two modules,we can not only mine the fine-grained discriminative information of different parts of the vehicle,but also learn the foreground features from the multi-scale spatial features.By doing this,the method can obtain discriminative information from the global and spatial local aspects respectively,and extract more robust feature representation.(2)Vehicle re-identification algorithm based on relation mining among images.In order to solve the problem of ignoring the structure information and correlative information between adjacent samples,a new algorithm of the vehicle re-identification algorithm based on relation mining among images is proposed.Firstly,the algorithm uses the graph attention mechanism to aggregate the vehicle nodes under different viewpoints and make full use of the contextual information contained in the cross-view vehicles,so as to extract feature representation that is robust to dramatic viewpoints changes.Secondly,the mask guided graph attention mechanism module is proposed.The negative sample center vector of each vehicle is used to eliminate the negative effects of visual appearance similarity caused by different vehicles under similar viewpoints,so as to narrow the positive sample pair in the feature space and push away the negative sample pair,which is useful to enhance the learning ability to discriminant features.Finally,the ablation experiments and comparison experiments with other mainstream algorithms are performed on two public datasets to verify the effectiveness of the method. |
