Person Re-identification Based On Multiple Granularities Resolution Networks

Person re-identification(ReID)is a sub-problem of image retrieval,which is intended to determine whether there are specific person in the pictures taken by different security cameras.In recent years,with the development of deep learning,person re-identification technology based on deep learning has made significant progress,and has been widely used in public security,security monitoring,smart city and many other fields.The main source of person re-identification image data is taken by the camera in the city.In the actual scene,due to the complexity of environmental factors,including low image resolution,person posture changes,person dress changes,light changes,occlusion and a series of problems,person re-identification task is more complex and challenging.At the same time,because the amount of person image data is very large,and most of them are unlabeled data,it costs a lot to assign labels manually.How to effectively use a large number of data is also a topic of general interest.In this paper,the corresponding methods are proposed to solve the problems of low resolution,poor expression of detail features and insufficient use of a large number of person images.Aiming at the problem of low image resolution and poor expression of detail features,this paper proposes a multi-granularity resolution network(MGRN)based on the method of deep learning,on the basis of traditional part-based multi-granularity network,we adds features about multi-resolution information,improves the ability of detail feature extraction.at the same time,attention mechanism gives full play to the advantage of multi-granularity,and makes the model learn the weight proportion of different granularity features.Finally,the person feature vector output in a structured way,which effectively improves the ability of feature expression of the model.The model has achieved the state-of-the-art of person re-identification task.To solve the problem that the previous models ignore the partial similarity of pedestrians in the task of cross domain person re-identification.This paper proposes a multi-granularity label method,which calculates the feature vector of each branch output from MGRN,obtains the "pseudo label" of the target domain by clustering and assigns it to the target domain dataset,to build multi-granularity labels with one primary identity and multiple part identity for each pedestrian image,and cover the similarity of the feature part.The model is trained iteratively in the target domain,which makes the model adaptive to the target domain dataset.