Vehicle Re-recognition With Attentional Mechanism

Vehicle recognition is an important part of smart city and a brand new technology in the field of video.Compared with pedestrian re-recognition,vehicles have very high similarity in appearance,and different camera positions lead to different pictures,which brings many challenges to vehicle re-recognition.This brings many challenges to vehicle re-identification.Therefore,it is more important to use neural network to mine fine-grained features with identification.Aiming at these problems,this paper studies vehicle re-recognition methods from three aspects: the construction of wavelet convolution network,multi-scale feature extraction of vehicle image and the attention mechanism of neural network.The specific work and innovations are as follows:(1)Proposed a convolutional network vehicle re-recognition algorithm based on the combination of wavelet feature and attention mechanism.In view of the advantages of image wavelet transform in multi-resolution and local analysis,a wavelet spatial attention module is designed based on the neural network model.In order to improve the identification ability of vehicle image features,the wavelet module is embedded in the attention network,and the wavelet transform is used to replace the pooling layer and serve as the lower sampling layer.In this way,the features of vehicle images input into the model after wavelet transform and convolution layer will contain different frequency and scale information,effectively improving the accuracy of the model.Through a lot of experiments on the vehicle weight recognition data set,the average accuracy is improved effectively.(2)A convolutional neural network model with multi-scale step fusion network is proposed for vehicle re-recognition.In this model,the global context information is first learned through different feature layers.Then,feature details of multi-scale layers are fused to improve the complementary ability of each feature at different scales.Combined with the wavelet attention mechanism model,the ladder network obtains more discernible features.Through experiments on Ve Ri data sets,the performance of the model is improved.(3)A multi-learning strategy training method is proposed,which employs cross entropy loss function and difficult sample ternary loss function to train the joint network model.Through comparative analysis on the open data set Ve Ri data and the Vehicle ID data set,good results are achieved in both Ve Ri and Vehicle ID.