Research On Vehicle Re-Identification Method Based On Attention Mechanism And Multi-Granularity Feature Learning

Vehicle re-identification refers to the technology of quickly and accurately identifying and locating target vehicles through feature extraction and matching of vehicle images in surveillance videos.It has important application value in the fields of smart city,intelligent transportation,automatic driving and criminal investigation.With the development of deep learning,vehicle re-identification technology can extract vehicle appearance features more accurately and in detail,thereby significantly improving the accuracy and reliability of vehicle recognition.However,since the vehicle images come from different cameras,the viewing angles,illumination and resolutions of the vehicles are different,which brings the challenges of small inter-class differences and large intraclass differences to vehicle re-identification.The key to accurately distinguishing different vehicles is to extract discriminative vehicle features.To this end,this paper constructs two novel and effective network models based on attention mechanism and multi-granularity features.(1)For the problem of large intra-class differences and small inter-class differences,this paper uses the attention mechanism to enhance the feature representation of vehicles,and proposes a three-branch network model based on global relational attention and multi-granularity feature learning.First,a global relational attention mechanism is designed,which models the average pairwise relation between a certain position/channel and all other positions/channels,obtains the global relation of the position/channel,and then derives the global importance of position/channel.On the one hand,this mechanism reduces the difficulty and computational complexity of attention learning;on the other hand,the global relation can robustly evaluate the importance of locations/channels,making attention learning more accurate.Based on this mechanism,this paper constructs spatial global attention and channel global attention modules,and adds them to two independent branches to extract effective global features from different dimensions.Second,to obtain the fine-grained features of the vehicle,this paper divides the vehicle feature map evenly into two parts by using a hard partition strategy according to the characteristics of the vehicle re-identification task.(2)Although the hard partition strategy can extract local features simply and effectively,using only local features without considering the correlation between vehicle parts cannot distinguish similar vehicles well.Because there are similar attributes in corresponding parts between vehicle images belonging to different identities.To solve this problem,this paper designs a cross-local interaction module,which restricts the self-attention calculation to parts through window mask to strengthen the association between parts.Robust local-level features that are more beneficial for identifying vehicle identities can be obtained by considering inter-part relationships.Secondly,the weight generation method of the channel global attention module is optimized,and the grouping aggregation mechanism and attention enhancement constraint are proposed to reduce the parameter amount of the module and further improve the network’s ability to recognize vehicles.Finally,this paper constructs a network model based on grouping aggregation mechanism and cross-local interaction.