| Text-based person Re-identification(Re ID)aims to retrieve corresponding pedestrian images from an image database using textual descriptions.This task has attracted widespread attention from researchers in the fields of smart cities and intelligent security.However,there are still some issues with the current research on cross-modal person Re ID algorithms that utilize text and images.For example,local-global cross-modal matching has been shown to perform well in cross-modal person re-identification algorithms,but there are still some problems with the partitioning of local features.Uniform segmentation without processing can lead to the loss of a lot of edge information,and segmentation after data preprocessing will increase additional computational complexity.In addition,due to the explosive growth of image and text data,there is a problem of feature dimensionality catastrophe,which makes high-dimensional feature data unable to balance retrieval performance and retrieval efficiency.To solve the above problems,this article proposes three different cross-modal person re-identification algorithms,including a multi-view feature fusion-based cross-modal person reidentification algorithm,a multi-branch cross-modal person re-identification algorithm based on attention hash coding,and a fast two-stage cross-modal person re-identification algorithm.The main work of this article is as follows:(1)Previous research has focused on local-global feature matching between images and text,but these studies have ignored some boundary information due to uniform segmentation.We propose a new cross-modal person re-identification model called Multi-branch Fusion Network to better map the features of both modalities into the same semantic domain.We first propose a multi-view module that uses a dual-layer dilated convolution to capture key image features.Then,we design a multi-branch module to eliminate boundary loss caused by uniform segmentation.Finally,a mixed matching method from low-dimensional to highdimensional and from local to global is adopted.A large number of ablation experiments and comparative experimental results show that the accuracy of our proposed Multi-branch Fusion Network model is significantly better than existing mainstream algorithms.(2)To address the problem of feature dimension explosion in cross-modal pedestrian re-identification,a channel attention-based hashing encoding algorithm was proposed.To solve the fine-grained retrieval problem in image-text hashing,a block mapping hashing learning method was proposed to effectively improve the interference between individual hash codes and mine local information of person features to enhance the matching degree between different modal hash codes.Then,the traditional uniform partition method in pedestrian reidentification algorithm was improved,and a channel attention-based hashing encoding method was designed to allocate more weight to the region of person features that deserve attention.Experimental results showed that our Hash-Re ID reduced the feature code dimension by 32 times compared to 2048-dimensional features,and the retrieval speed was increased by nearly 8 times.(3)To balance the performance and efficiency of the algorithm,highdimensional features and hashing coding features were jointly applied in a twostage retrieval strategy.In addition,a fast cross-modal pedestrian re-identification algorithm was designed by combining Res Net-50 and BERT models.Experimental results showed that compared to other text-based pedestrian reidentification algorithms,the model had an absolute time advantage,with the feature dimension for final retrieval reduced by 32 times,and the retrieval precision of Rank-1,Rank-5,and Rank-10 improved by 2.17%,2%,and 1.99%,respectively,compared to the current state-of-the-art algorithms. |
