Vehicle Re-identification Model Based On Improved ResNet Network

With the continuous improvement of people's material life,the number of vehicles has also increased year by year,but traffic safety issues have also become more frequent.Every year,traffic problems such as illegal cases and law and order management continue to increase,such as license plate blocking,license plate decking,and vehicle escape.These problems make people pay more and more attention to the problem of traffic safety.Based on the above background,to effectively alleviate the contradiction between the current situation of traffic management and people's safety requirements,it is necessary to quickly find the target vehicle in the massive vehicle database,that is,vehicle re-identification.Aiming at the research content of this topic,we have studied the feature extraction method that can better express the characteristics of vehicles,and can effectively improve the accuracy of vehicle re-identification.The main research is as follows:(1)Summarize the research progress of existing vehicle Re-ID work.According to different methods,from traditional visual features to deep learning features,as well as local features,representation learning,metric learning,and GAN network methods,The disadvantages are compared in multiple dimensions,and the commonly used normalization algorithms are summarized.(2)In this work,we study a deep neural network based on a normalized vehicle Re-ID and propose a strong basic model.In our research work,instance-batch normalization is used in the shallow network layer by combining instance normalization and batch normalization.In IBN,using IN eliminates appearance differences,while using BN is essential for extracting content information.In the deep network layer,batch normalization is added after the global pooling layer.Because the features of low-level and middle-level layers represent the appearance information,and the features of the high-level layers are the semantic information.IN and BN are introduced in the shallow layer,while only BN is added in the deep layer.By properly adding normalization to the ResNet network in this way,the network can reduce intra-class differences and improve performance.(3)Due to the large intra-class gap caused by different noises,colors,offsets,brightness,or styles,we improved the ResNet network model to eliminate the impact of intraclass differences on the model,and through reasonable design.The model makes it possible to retain the discriminative power of the pre-trained model and prevent overfitting to make the model performance better on the new data set.After researching different normalization methods,a reasonable normalization scheme is proposed,which provides a powerful benchmark model for vehicle Re-ID,and achieves the best on VeRi-776,VehicleID and VERI-Wild datasets.The model can solve the problem of too large intra-class gaps,and the model has good robustness,anti-interference,and easy migration.