Research On Vehicle Re-identificaiton Algorithm Based On Deep Learning

Vehicle Re-identificaiton(Re-Id)task is to search the image of a specified vehicle in an image database containing multiple cameras.It has universal applications in video surveillance,intelligent transportation,and city computing.It is a frontier and important topic in the field of computer vision research.The field of visual research is a cutting-edge and important subject.At present,the existing methods for vehicle re-identification mainly focus on deep learning models based on appearance features,but it is difficult to obtain the best results due to slight differences in different states of the same vehicle and similar appearances of different vehicles.In view of the above problems,this article conducts in-depth research on the appearance-based vehicle re-recognition deep learning model Rep Net.The main research work and results include:(1)A migration pre-training method based on GAN imbalanced samples is constructed.Due to the incomplete sample of different data sets and the imbalance of the number of different vehicle samples in the same data set,the training effect of the deep learning model is not ideal,and the promotion performance becomes poor.Use the GAN network to learn the samples,generate new samples,pre-train Rep Net on the new sample set,and then use this model to perform migration training on the original data set.The experimental results show that compared with the original model,the GAN-based imbalanced sample migration pre-training method has improved the detection effect to a certain extent.(2)A vehicle re-recognition algorithm based on metric learning and multi-channel model is constructed.This article uses Res Net instead of VGG network on the basis of Rep Net,deepens the network layers,and extracts richer vehicle fine-grained features;replaces the fully connected layer with a 1 * 1 convolution operation to achieve the input of pictures of any size;The two channels of vehicle color are combined with coarse-grained appearance and fine-grained appearance to perform vehicle feature learning.ARC Face Loss is used instead of Triplet Loss to convert the similarity measure to spherical space,increasing the distance between classes,reducing the distance between classes,and enhancing the alignment Classification capability of similar vehicles;reorder the sorted list output by the model based on the mean of Euclidean distance to optimize the retrieval results.Experiments show that compared with the classic Rep Net algorithm,and can achieve good detection results.Compared with some other classic algorithms,our algorithm has achieved the best recognition performance.