Research On Vehicle Re-identification Method Based On Multi-scale Joint Learning

With the development of urbanization,vehicle as an important means of transportation has become more and more important,but it also brings challenges to traffic management.In public security,the need to search for specific vehicles has also become an urgent problem to be solved,so the task of vehicle recognition has also attracted more and more attention.The simplest way to recognize vehicles is license plate recognition.However,due to the scene,angle,light and the appearance of license plate vehicles,there will be unrecognizable situations.So for this problem,we do not use the vehicle information to solve the problem of vehicle recognition,only rely on the information captured by the camera to identify.Similar to other target recognition,different vehicle models,changing lighting and complex environment have a serious impact on the effectiveness of vehicle recognition.Therefore,we propose a vehicle recognition scheme based on deep learning for vehicles.This paper mainly includes the following contents.1)Firstly,we introduce the research background and significance of vehicles,and make a corresponding investigation and analysis of the previous vehicle weight recognition schemes,and summarize their advantages and disadvantages as well as what else we need to solve.At the same time,the network model proposed by us is also introduced.2)To solve the problem of vehicle recognition,we use triple loss to train convolutional neural network,which automatically extracts features.In training,we use triple(anchor,positive sample,negative sample)to capture the relative similarity between them to learn representative features.However,due to the weak constraint of the traditional triple loss training,the expected results have not been achieved.We propose improvements to train from three aspects:first,for the weak constraint of traditional triple loss,we propose batch difficult triple loss,which not only solves the problem that many triples become "useless" triples with the continuous deepening of training,but also solves the problem of training time-consuming.Second,we use more powerful cross-entropy loss and difficult triple loss joint training to strengthen its constraints.Thirdly,in order to reduce the generalization ability and adaptability of the model due to overfitting,we add label smoothing regularization(LSR).We evaluate the proposed method on the benchmark data set,and the comprehensive experimental results show that the proposed method has good performance compared with the existing technology.3)For similar vehicles,we propose a multi-scale fusion vehicle re-identification network to achieve the task of vehicle recognition.First of all,our method focuses on learning strong feature representation to distinguish similar vehicles,training data with multi-scale convolution neural network fusion framework,and using different expansion ratio in each scale to capture different text information.After training with the combination of deep and shallow network,the shallow network can produce less invariance and low-level image features,and the deep network provides high semantic information.Finally,it is fused in the softmax layer to effectively avoid the lack of detailed information.Finally,the experimental results are rearranged by k-order reciprocal.One is the final distance,which is weighted by two tributaries,one is the original Mahalanobis distance,and the other is the Diskal distance calculated from the extracted k-order reciprocal.It not only makes up for the missing query vehicles,but also reduces the interference of negative samples.Our experimental results show the effectiveness of the proposed method.