Clothing Parsing Based On Height Class Distribution And Feature Alignment

With the development of e-commerce and the rise of various online shopping platforms,artificial intelligence technologies such as deep learning are increasingly used in scenarios such as image retrieval,clothing recommendation and virtual fitting.As the key technologies of these application scenarios,the task of clothing parsing is to use fine-grained semantics to segment clothing images into multiple parts and assign semantic labels of clothing type or background to each pixel,so as to provide accurate semantic segmentation results for the application of clothing image recommendation,processing and other technologies.Since there are many kinds of clothing images,complex images and different clothing parts are easy to block each other,how to achieve accurate clothing analysis is an important research task in clothing image processing.In this paper,an improved clothing analysis algorithm based on height class distribution and feature alignment is proposed to solve the problem of feature misalignment and improve the ability of spatial modeling.The research content of this paper mainly includes:(1)In view of the characteristic that all kinds of clothing in clothing images have relatively fixed Height and position Distribution,this paper designed a Height Class Distribution Guided Module(HCDGM)to guide the learning of height and position information in clothing images.This module simplifies the complex position information of clothing into the vertical position information of each clothing,and guides the parsing network to use the internal position distribution information of clothing to guide the segmentation,and improves the ability of the model to extract the height distribution information.(2)Aiming at the dislocation of some edge information that may be caused by the up-down sampling process of the clothing parsing pyramid structure,this paper designed the Clothing Feature Aligned Module(CFAM)to align the dislocation edge feature information.In this module,the acquired features are screened to a certain extent,and the sampling position in the convolution kernel is adjusted by a set of learnable feature mappings so as to adapt the misaligned edge information.(3)In this paper,Shifted Patch Module(SPM)is designed to improve the local induction ability of backbone network,so as to solve the problem that transformer model used for computer vision will focus more on global information and lack of access to local information,and play a certain role in data enhancement.Based on Swin Transformer V2,the block embedding layer is modified to improve the performance of the model on a small data set.In this paper,validation experiments are carried out on CFPD data sets,and the results show that the parsing accuracy of each data set is improved to some extent compared with CCNet,Deeplabv3,PSPNet and other algorithms.Among them,Swin Transformer network for clothing parsing based on height class distribution guidance and feature alignment reaches 93.93% PA and 54.66% m Io U in CFPD dataset.On this basis,after adding block transfer module and modifying the backbone network,the segmentation accuracy is improved by 0.57% in m Io U when the backbone network with the same parameter size is used.Experiments show that the proposed method can accurately parsing clothing objects of different scales and eliminate some abnormal clothing combinations,which has good practical application value.