Vehicle Re-identification Based On Super Resolution And Pyramid Convolution Residual Network

Vehicle re-identification(Vehicle Re-ID)aims at retrieving and tracking the specified target vehicle with multiple other cameras,which can provide help in checking violations and catching fugitives,but there are still the following problems that need to be solved urgently.First,the existing collected Vehicle Re-ID data often have low resolution and blur in local areas,so that the recognition algorithm cannot accurately extract subtle representations.In addition,small features are easy to cause the disappearance of features under the operation of large convolution kernels,which makes the model unable to capture and learn subtle features,resulting in inaccurate judgment of vehicles.Second,compared with the convolutional neural network,the advanced Vision Transformer requires significantly less computing resources during training,but the existing Vision Transformer-based Vehicle Re-ID method adopts the absolute position encoding method,which cannot obtain the information between the local positions of the vehicle.Relative meaning and no constraint relationship between positions make the network unable to robustly learn the connection information between vehicle components.Aiming at problem one,this paper organically combines super resolution and Vehicle Re-ID algorithm,and proposes a Vehicle Re-ID method based on super resolution and pyramid convolution residual network.Firstly,by learning image feature fusion and transferring low-frequency and high-frequency feature information through generative adversarial networks,low-resolution images are converted into super-resolution images with richer texture and color information,thereby providing more indispensable details.Then,the multi-scale residual features of the input images are extracted through the pyramid convolution layer with multiple receptive field sizes,which can capture information on different scales.Finally,the feature fusion operation makes the fused features have more discrimination advantages and greatly improves the robustness of image features.In order to verify the effectiveness of the method,experiments are carried out on Ve Ri-776 and Vehicle ID datasets.The experimental results show that the method proposed in this paper effectively improves the data information of vehicle images,accurately distinguishes the subtle differences between different vehicles of the same model,and is superior to other methods in terms of recognition accuracy.Aiming at problem two,this paper proposes a Vehicle Re-ID method based on Vision Transformer with super resolution and relative position coding.The first is to use super resolution operation to provide more indispensable details.Then the relative position encoding is used,which can improve the model’s ability to understand the semantic association information between the various slices of the vehicle image,so that the network can robustly learn the connection information between vehicle components.Secondly,the side information embedding module is introduced to process non visual information such as camera information and viewing angle of vehicle images.Finally,the puzzle module is used to introduce additional disturbances in the training and improve the robustness and accuracy of the Vehicle Re-ID model.In order to verify the effectiveness of the method,experiments are carried out on Ve Ri-776 and Vehicle ID datasets.The experimental results show that the method proposed in this paper can effectively improve the data information of the vehicle images and the Vehicle Re-ID model’s ability to understand the semantic correlation information between each slice of the vehicle image,and is better than other methods in terms of recognition accuracy.