An Attribute Recognition Method Based on ResNeSt and Person Online Hard Example Mining

In pedestrian attribute recognition based on deep learning, severe class imbalance between positive and negative samples degrades recognition accuracy. This study proposes an attribute recognition method that integrates ResNeSt and Person Online Hard Example Mining (POHEM). The method employs a ResNeSt-50 backbone and a Channel Attention Module (CAM) to extract salient attention features. It enhances model performance through label smoothing, the POHEM strategy, and joint multiattribute training. Experiments on the RAP and PA100K datasets demonstrate that the proposed approach outperforms existing pedestrian attribute recognition models in overall performance, exhibiting strong applicability and competitiveness.