Research On Person Re-identification Algorithm Based On Attention Mechanism

As the public's demand for safety continues to grow,person re-identification plays an increasingly important role in the field of computer vision,and is widely used in many real scenarios such as smart security,video surveillance,military surveillance,etc.Person re-identification can be regarded as a retrieval problem,that is,given a candidate person image,retrieve the same person image from the gallery of different monitoring equipment.In recent years,the rise of deep learning has brought more attention to person reidentification and has achieved fruitful results.However,due to the huge changes in illumination,occlusion,resolution,background and other factors,as well as the lack of person tags,the task of person re-identification still faces great challenges.Therefore,the research on the performance enhancement of person re-identification model and the enhancement of model generalization ability has important application value.This thesis focuses on the problem of optimization of feature extraction process and the problem of unsupervised domain adaptation,and proposes different methods to solve these problems.The main work and innovations are as follows:In the existing deep learning-based person re-identification methods,the extraction of good features is still a key step.However,the features learned by convolutional neural networks without guidance are often simple and redundant.The thesis introduces the second-order statistic covariance based on the non-local operation to capture the correlation information between different areas of the person image,and improve the characterization ability of person characteristics.Aiming at the problem of feature redundancy,the thesis also uses the expectation maximization algorithm to reconstruct features with low rank,and reconstructs attention features with low redundancy characteristics through multiple iterations.The thesis evaluates the proposed method on three popular benchmark data sets: Market-1501,Duke MTMC,and CUHK03,and the experimental results prove the superiority of this method.The attention model-based method proposed in the thesis has achieved excellent performance,but the effect is not good when performing unsupervised training on unlabeled data.To solve this problem,the thesis proposes an unsupervised person reidentification method based on the asynchronous mutual teaching framework based on the theory of OPTICS clustering algorithm.First,the thesis uses an optimized clustering algorithm to divide the unlabeled data into clustered values and outliers,and input them into two networks for mutual training and learning,so as to reduce the impact of noise.Second,in order to maximize the use of difficult samples in outliers to improve the model's ability to recognize difficult samples,the thesis introduces a cache module to store the iconic features in the clustering results,and proposes a joint loss function to constrain the input data for the training process.Experimental results show that the method mentioned can fully mine and use clustering difficult samples,effectively improving the performance of the model.