Clothing Attribute Recognition And Application Research Based On Anchor-free Detection Framework

In the past few decades,with the rapid development of e-commerce platforms,a large number of clothing images have appeared on social media such as the Internet or mobile applications.Due to the diversity of colors,textures,styles and categories of clothes,it will be more difficult for people to choose the right clothes when shopping online.Clothing attribute recognition,as a key technology for users to describe clothing features,is of great importance to solve this problem.To handle this problem,most current methods are firstly to detect multiple clothes,then to cut out the clothes,and feed them back to an additional network for clothes attribute recognition.This two-stage approach is time-and resource-intensive;on the other hand,the one-stage approach can provide an effective and efficient solution by integrating clothes detection and attribute recognition into an end-to-end framework.However,one-stage methods tend to use anchor-based detectors,which will lead to high sensitivity to hyperparameters and high computational complexity for dense anchor boxes.In addition,since the clothes detection branch and the attribute recognition branch require different optimized features,they may also confront the problem of optimization contradiction during the training process.The handle the above problem,we aim to develop new effective method for clothes attribution detection.The main contributions of the proposed work can be summaried as follows:1.To address the issues of two-stage methods and the limitations of anchor-based methods,we develop a clothing attribute recognition network based on an anchor-free detection framework.It is a one-stage end-to-end anchor-free framework containing an additional branch for joint clothes detection and attribute prediction tasks.The two task branches in the network share the characteristics of the backbone network and perform calculations in parallel,which greatly improves the recognition rate.2.To further improve the prediction accuracy,we introduce an attribute grouping attention module in the recognition branch.It can adaptively integrate mutual exclusion and association between attribute types into feature learning,and a large number of experimental results verify the effectiveness of the proposed module.In addition,in the learning process of the recognition branch,we adopt a loss function suitable for multi-label classification tasks to further improve the learning effect of the network.3.To handle the optimization contradiction in the two task branches,we propose a backbone feature decoupling module.It encodes the backbone features into pixel-level dense query features,and decodes them into output features via a deformable transformer,which are fed into the clothing detection and attribute recognition branches,respectively.In this way,the features of the detection and identification branches can be decoupled,and the optimization contradiction can be naturally resolved.4.Finally,we explore the problem of similarity matching between clothing attribute text and images.We built a graph-text matching network based on the CLIP model.The network involves two modalities of image and text,it encodes images into feature vectors,and encodes text descriptions into sentence vectors.We employ a contrastive learning approach to map images and text into the same semantic space,and finally apply it to the task of attribute text-based image retrieval for clothes.Relevant experiments show that the image-text matching network we constructed has a certain matching and discrimination ability for related attribute texts.