Clothing Image Classification And Detection Based On Swin Transformer

With the rapid advancement of Internet technology in both personal and professional contexts,online shopping has gradually become the preferred choice for consumers when it comes to purchasing clothing,food,housing,and transportation.When it comes to shopping for clothing online,consumers often turn to major e-commerce websites such as Taobao and JD.com.However,due to the vast array of clothing styles and product categories available on these websites,classifying various product pictures has become a crucial aspect of the online shopping experience.Traditional image processing methods are no longer sufficient to meet these needs,and in recent years,image classification and detection tasks have become key issues in deep learning research.As a result,image classification and detection technologies based on deep learning have been widely implemented in various fields.During the process of clothing image classification,there are challenges such as a wide range of images,low feature richness,single feature scale,and complex image background environments that can interfere with clothing classification.This paper addresses these issues by studying the classification and detection of clothing images using the Swin Transformer network,which has both theoretical significance and practical applications.(1)The information enrichment characteristics of the frequency domain are studied.According to the information of clothing images,the Fourier transform is used to extract the frequency domain information of the image,and a certain threshold radius is taken to separate the high and low frequency information of the frequency domain information,and the high and low frequency information images under different threshold radii are obtained,and then the processed clothing image is mapped back to the spatial domain,and the experiments show that the frequency domain features of rich images have a certain effect on improving the accuracy.(2)A weakly supervised clothing image classification method is proposed.In the case of only the category information of the image,but does not contain any other location or label information,a weak supervision mechanism is introduced for the classification of clothing images.Through experimental training and verification on the test set,the results of clothing image classification based on weak supervision are obtained.(3)Clothing detection of multiple targets in a single image using strongly supervised object detection.Using the position labeling of clothing,using Swin Transformer as the backbone network,and adapting the Faster RCNN model to generate suitable anchor frames according to the dataset size,using Soft NMS for detection frame filtering,and ROI Align to reduce errors.Then training experiments and test work are carried out to obtain the results of object detection experiments.In this paper,aiming at the problems of low feature abundance and single feature scale in clothing image classification,as well as the interference caused by complex image background environment on clothing classification and the difficulty of multi-target detection,experimental verification is carried out on the Deep Fashion subset,and the experimental results show that the weakly supervised clothing image classification and clothing image detection proposed in this paper can achieve good experimental results.