Research On Unsupervised Person Re-identification Based On Multi-granularity And Quadratic Clustering

Person Re-Identification is a technology for matching and retrieving images of the same person captured by different cameras.It is one of the core technologies of intelligent security monitoring and has a wide range of application requirements in the fields of video surveillance,security protection,criminal investigation and public security.However,the supervised Person Re-Identification algorithm requires a huge Datasets of labeled images to obtain the correspondence between images and persons.In order to save the cost of manual labeling and facilitate the rapid deployment of the model in different scenarios,research on unsupervised Person Re-Identification method has important practical significance.The purpose of unsupervised Person Re-Identification is to match the query image of persons with the Datasets images without any label information.Compared with supervised Person Re-Identification,unsupervised Person Re-Identification does not require any labeled Image information can train a model with generalization ability,which is more challenging.However,existing unsupervised Person Re-Identification do not take into account issues such as differences in person images style and pseudo-label noise.Therefore,this paper proposes two different algorithms to improve and solve the above problems.The specific research content is as follows:(1)Considering the impact of objective factors such as hardware differences and lighting changes,the image style of the same person captured by different cameras is prone to large contrast,which in turn makes the image features too different,and the contrast loss function continues to increase,making it difficult for the network model to Converge further.Therefore,this paper proposes Unsupervised Person Re-identification Based on Quadratic Clustering.The method mainly includes a global quadratic clustering module and a local supervised learning module.Used to solve the problem of inconsistent imaging styles and imbalanced number of cluster instances in the memory dictionary for the same person under different camera perspectives.In order to illustrate the superiority of this method in unsupervised Person Re-Identification,a large number of experiments were carried out in this paper.The results show that,compared with other advanced unsupervised Person Re-Identification algorithms,this method significantly improves the accuracy of unsupervised Person Re-Identification,The m AP of 81.2%,68.4%,31.1%,88.3% and 38.4%were achieved on the Market-1501,Duke MTMC-Re ID,MSMT17,Person X and Ve Ri-776 Datasets,which can better solve unsupervised Person Re-Identification problem.(2)The above method does not further improve the clustering pseudo-label noise and image feature expression ability.Based on this,this paper further improves the above method and proposes Unsupervised Person Re-identification Based on Multi-Granularity Feature Fusion.This method proposes a dual network model to suppress pseudo-label noise.On the other hand,this method performs multi granularity segmentation,cropping,and fusion of pedestrian images,allowing the model to learn more robust features.Subsequent experiments show that the method performs better on multiple public Datasets.On the Market1501,MSMT17,Person X and Ve Ri-776 Datasets,the m AP were 83.4%,32.1%,89.1% and 40.0%,significantly improving the performance of unsupervised Person Re-Identification algorithms.