| In recent years,deep learning has been widely developed in the field of artificial intelligence,especially Convolutional Neural Networks(CNN).Person re-identification,as a hot and key branch in the computer vision field today,by inputting the image or video data collected under different cameras,which can realize the similarity retrieval of specific pedestrians.Person re-identification is regarded as a subtask of the image retrieval,combining it with CNN will greatly improve the performance of the re-identification algorithm through large-scale data training.In the current research,together with the latest intelligent technology,person re-identification task can be widely used in specific scenarios,such as unmanned supermarket supervision,community safety protection and child search,etc.Generally,person re-identification technology based on convolutional neural network is mainly studied from two aspects: feature extraction and similarity measurement.Although the technology of person re-identification based on deep learning is developing rapidly,there are still many difficulties in practical applications: 1)With the introduction of deeper and wider networks,the detail information in low level is often lost as the convolution layer deepens,which makes the algorithm lack of robustness;2)Suffer from background redundancy,illumination variation,and occlusion deformation,which makes pedestrian images are not aligned.Consequently,the differences between pedestrian images are very large and difficult to distinguish;3)The fusion method of combining local features lacks the analysis of the relevance and consistency between local regions;4)The single metric loss function cannot solve the problems of intra-class difference and inter-class similarity very well,and cannot obtain very good results for complex distributed image data.According to the questions mentioned above,a new person re-identification method is proposed based on convolutional neural network in this paper.The main research works are as follows: 1)This paper used the deep network Resnet50 as the framework,by deepening the relationship between high and low level modules to enhance the robustness of features and improve the performance of the model;2)Designing the network structure to extract robust feature is an important part of person re-identification technology.This paper used the deep network Res Net50 and embedded the spatial transformation structure to extract semantically consistent local features,which advanced the robustness of obtaining expressive features.By solving the problem of local spatial semantic feature inconsistency,this paper accurately expressed the main characteristics of the target,and achieved pedestrian alignment;3)In order to explore the correlation between local features,this paper designed a strong feature fusion module based on the aligned local features to make full use of semantic information,which enhanced the discriminative power of the network;4)To unleash the discrimination ability of the strong representations of this network architecture,this paper used classification loss and triplet loss to train the network,and Ranking Matrix(RM)method was proposed to select local triplet samples.In order to solve the problem of large intra-class differences and small differences between inter-class,we designed a local triple loss to improve accuracy,and the experiments proved that the result is very good.Finally,we compare our method with the current state-of-the-art person re-identification approaches.The experiment verified that our method have better results on person re-identification task. |
