A novel Swin Transformer based on height class distribution and feature alignment for clothing parsing

Zhexin Huang, Shaoyu Wang, Lan Li, Jingyi Fan, Xiujin Shi

Abstract: The task of clothing parsing is to assign the corresponding clothing category label to each pixel in the clothing image. In this paper, an improved Swin Transformer clothing parsing method fusing height class distribution and feature alignment is proposed to solve the problem that existing clothing parsing algorithms rarely consider position and feature alignment, which leads to unsatisfactory segmentation accuracy. Specifically, we propose a Height Class Distribution Guided Module (HCDGM) to strengthen the guiding role of position information in the training process of clothing praising task. To further improve the parsing accuracy, the feature information of adjacent depth is aligned by Clothing Feature Aligned Module (CFAM). The proposed Swin Transformer based on HCDGM and CFAM can effectively use the height position features of clothing images and correct the information misalignment between different scale features. Experiments results show that our method achieves better results (54.66% of mIoU and 93.93% of PA) on the CFPD dataset compared with other advanced methods.