| Person Re-identification(Re ID),also known as pedestrian re-identification,refers to the search for the same pedestrian under different cameras,that is,given the pedestrian image under a monitoring device,the image of the person in the row under the cross-device is retrieved.It has high application value in intelligent security,intelligent person seeking system,intelligent business,large shopping malls and unmanned supermarkets,photo album clustering,family robot and other fields.In the deep learning method of convolutional neural network(CNN),the slicing method of image-based person re-reidentification has overlapping problems between slicing due to CNN’s receptive field,which leads to cross-aliasing phenomenon of features,limiting the number of blocks and affecting the classification performance of the model.Therefore,we propose CET and CIT hybrid network model algorithm on the basis of CNN and Transformer deep network framework,aiming at the limited number of blocks caused by the overlapping of the receptive field,and the phenomenon of cross-aliasing and unalignment of block features in person re-identification slicing method.The main research works are as follows:1.In order to alleviate the problem of the limited number of slices caused by the overlapping of block feature sensing fields in the horizontal slice method of person re-identification,a network structure CNN with INOUT＿Transformer(CIT)based on Transformer is proposed.First of all,the input image is obtained through CNN feature extraction network,and then the feature map is divided into blocks,and each piece of feature is further cut into pixel-level token vectors.After flattening each pixel level token vector and adding the position encoding and global token vector,they are input into the Transformer IN encoder.Then,the same operation is performed for each slice feature,and the encoding of each slice feature is taken to output the global token vector.The obtained global token vector is further added with classified token and position encoding and then input to the Transformer OUT encoder to obtain the final encoder output.Finally,after taking the classification token vector and adding the fully connected layer,the pedestrian is classified by softmax and cross entropy loss.Experiment on Market-1501 and Duke MTMC-re ID datasets show that the proposed method based on transformer can extract features more fine-grained,and further improve the number of slices and classification accuracy by utilizing Transformer’s global control ability.2.In addition,in order to alleviate cross-aliasing and feature alignment problems,we propose CNN with Embedding Transformers(CET)person re-identification network structure combining the respective advantages of CNN and Transformer.In CET,the Residual Transformers(RT)structure is embedded In the backbone network of CNN to obtain a feature extractor,Transformers In CNN(TIC).Transformer is good at modelling the relevance of global information.The feature aliasing phenomenon is improved.Secondly,a Feature Fuse with Learnable Vector(FFLV)structure is added to the output of Transformer at the end of the network to Fuse the output token vectors.Two Branch Loss(TBL)structure was designed to balance the two different fusion strategies.Finally,self-attention mechanism in Transformer is used for indirect alignment of human body parts,which solves the alignment problem caused by inaccurate detection frame.Experimental results show that CET network structure achieves better performance than Part-based Convolutional Baseline(PCB)and some other block slicing methods. |
