Lifelong person re-identification method and system based on decoupling of reinforcement of visual text attributes

Abstract

The application discloses a lifelong pedestrian re-identification method and system based on visual text attribute decoupling reinforcement, and belongs to the technical field of image processing. In order to solve the problems of knowledge forgetting and insufficient use of fine-grained attributes in the multi-domain continuous learning process of lifelong pedestrian re-identification, a pedestrian image training set is constructed, global and local text attribute features and visual attribute features are extracted, cross-modal alignment is performed in combination with a visual model and attribute query, and continuous learning and knowledge retention of the model are realized by optimizing the total loss. The application enhances the knowledge retention ability and representation discrimination ability of the model by jointly modeling the attributes of pedestrian images in the current domain and the historical domain, thereby improving the stability and precision of the lifelong learning system and the accuracy and adaptability of pedestrian re-identification.

Images