Research On Clothing Style Classification Based On Image Mask

With the rapid development of the clothing e-commerce industry,there are more and more clothing images on the internet,and businesses and consumers urgently need more and more intelligent solutions to meet the various tasks related to clothing.As a fundamental task in the field,clothing classification tasks play an important role that cannot be ignored.This article mainly explores the problem of how to classify clothing styles of characters in real environments.However,there are extremely complex feature information and difficult to clarify correlations between real environments and clothing images.How to extract the most recognizable features from images and accurately complete the task of clothing style classification has become the key of research.Clothing image classification is a fundamental task in tasks such as clothing retrieval,clothing recommendation,virtual fitting,and fashion trend prediction.Therefore,studying clothing classification tasks in real environments has important research significance and commercial value.The main work of this article is in the following two aspects:(1)Aiming at the problem that the general segmentation network is not accurate enough in the segmentation of clothing image elements in real environments,a clothing segmentation method based on clothing texture and semantic decoding module is proposed.In order to achieve higher accuracy in garment segmentation,firstly,image texture features are extracted as separate channels for stacking,and attention mechanisms are used to enhance texture boundary information;Then,the backbone network uses Transformer structure to extract clothing feature information.Due to its unique self attention mechanism,it focuses on the association information between pixels,strengthening the connection of clothing context information;Finally,during the processing of the decoding module,the accuracy of the decoder’s recognition of clothing boundaries is enhanced through the differential processing of different levels of features and the hybrid optimization of the attention mechanism.Through the analysis of a large number of experimental results conducted on LIP and ATR datasets,it can be seen that the proposed clothing segmentation algorithm based on the clothing texture and semantic decoding module improves the average pixel accuracy by about 2%,and the average intersection to parallel ratio by about 1.5%,effectively improving the performance of the clothing segmentation algorithm,enabling accurate acquisition of mask images of clothing,meeting task requirements.(2)Aiming at the problems of complex background information,numerous clothing styles,and inaccurate classification in real environments,a clothing style classification method based on image style semantics and self attention was proposed.This method implements the task of classifying character clothing styles through a two-step strategy.Firstly,the mask information of character clothing images is obtained through a segmentation network,the shape of the mask is optimized using a clothing style semantic generation module,and a clothing style semantic map is obtained by comparing it with the original image.Then,the semantic filling module is used to fill and restore the occluded clothing pixels,so that the clothing style semantic map has a more complete clothing shape,Finally,a multi scale self attention classifier is used to obtain clothing style classification results.This article constructs a clothing style classification dataset for research on classification tasks based on clothing masks.Through the analysis of experimental results,it can be seen that the two-step strategy method proposed in this paper improves the accuracy by about 15% compared to the pure classification network structure,and uses the network structure designed in this article to improve the benchmark network effect by about 20%.The experimental results prove the superiority of the method proposed in this article in clothing style classification tasks.In general,in order to complete the task of classifying the clothing styles of people in the real environment,this paper has done a detailed and in-depth study in the task of clothing segmentation and clothing classification,and effectively improved the experimental effect by reducing the interference of irrelevant information,strengthening the correlation of core information,and improving the strength of feature information representation,which has been verified on a variety of data sets.It provides reference scheme and technical support for the research of subsequent tasks and even the implementation of industrial applications.