Research On Person Re-identification Method Based On Deep Learning

Person re-identification(Re ID)is a technology of recognising and retrieving specific person images across cameras and scenes,and it is widely used in various fields such as security surveillance,case investigation and smart cities.In recent years,the rapid development of deep learning technology has greatly improved the performance of person re-identification algorithms.However,the research of person re-identification technology is still very challenging due to camera shooting angle,light variation,occlusion,and small sample size.Based on deep learning technology,this paper conducts research from two aspects: model design and data augmentation,which builds and trains a person re-identification model.The main research contents of this paper are as follows:(1)Research on person re-identification method based on global featuresIn order to address the problem that the model features cannot fully represent the person information due to illumination and pose changes in person re-identification,a person reidentification method based on global features is proposed.Firstly,the improved backbone network R-Res Net50 is used to extract person image features;Secondly,the feature layers at different scales of R-Res Net50 are extracted to embed the attention mechanism DANet(Dual Attention Network),so that the model focuses more on the key information on the person;Finally,multi-scale feature fusion is carried out on the extracted key features to achieve complementary advantages among features,and the network model is trained using a multi-loss function strategy combining cross-entropy loss,Tri Hard loss and center loss.The experimental results show that the person features extracted by this method are more discriminative and this method improves the accuracy of person re-identification.(2)Research on person re-identification method based on global features and local featuresIn order to address the problem that global features tend to ignore the fine-grained information of person images,a person re-identification method based on global features and local features is proposed,which divides the network structure into three branches.Firstly,the global features of person images are extracted using a global branch combining DANet with multi-scale feature fusion.Secondly,a feature chunking and reorganisation principle is investigated to horizontally segment the feature map at a ratio of 1:1 and 3:1 respectively,and local branch 1 and local branch 2 are designed based on this principle to obtain local features with rich detail information.Finally,the global branch uses the multi-loss function strategy,and the local branch uses the cross-entropy loss function and the center loss function to train the network model.The experimental results show that the method can effectively extract the detailed features of person images and further improve the recognition accuracy.(3)Research on person re-identification method based on data augmentationIn order to address the problem that low recognition confidence and over-fitting of the model due to the small scale and low diversity of the person re-identification dataset,a person reidentification method based on data augmentation is proposed.An improved Cycle-Consistent Adversarial Networks is designed to generate person images by simulating the distribution of sample data,and to complete the expansion of the person re-identification dataset.The expanded dataset is used to train the person re-identification model based on global features and local features.The experimental results show that the expanded dataset with improved CycleConsistent Adversarial Networks not only avoids the problem of overfitting due to insufficient samples for deep learning,but also effectively improves the performance of person reidentification.