Attention Mechanism And Convolutional Network-Based Models For Clothing Classification Method

The clothing classification task aims to extract discriminative features of clothing images and predict their class probabilities through the classifier.With the rapid development of e-commerce platforms,the clothing classification method based on computer vision is widely used in clothing retrieval,clothing fashion prediction,clothing recognition,clothing recommendation,and other fields.In recent years,with the development of deep learning technology and the introduction of convolutional networks,the performance of clothing classification has been greatly improved.Proceeding from the theoretical research of clothing classification,based on the development and research status of clothing classification,this paper mainly does the following three aspects:(1)Firstly,the research on clothing classification was summarized from the traditional feature extraction method and deep learning method according to the feature extraction method.Secondly,six mainstream public clothing datasets are introduced,and three commonly used clothing datasets are selected to analyze and compare some typical clothing classification methods.Finally,the research status and existing problems in the field of clothing classification are discussed,and the future research direction of clothing classification has prospected.(2)Aiming at the problem that the existing clothing classification methods are sensitive to background noise and lack the ability to express multi-scale features,a multi-scale deep convolutional network model(MCA-Inception)combined with an attention mechanism is proposed.Inspired by the Inception V3 network structure,convolution kernels of different scales are added to each Inception module to achieve regional perception under the multiscale receptive field to enrich the contextual details of features.In addition,the CBAM attention module is embedded in the improved backbone network to suppress the interference of noisy information such as chaotic background and enhance the ability of the model to express important information.The experimental results on Deep Fashion and ACS datasets show that the classification accuracy of the proposed method is improved by 4.02%and 4.82% respectively compared with the benchmark network,which proves that the proposed method can effectively improve the classification accuracy of the model.(3)Aiming at the cumbersome process of extracting key features in clothing classification,the amount of calculation is large,and the spatial neighborhood correlation is easily ignored,a clothing classification model based on the COT attention residual network is proposed in this paper,which aims to improve the accuracy of the classification and reduce the parameters and computation of the network model.This model uses Res Net101 as the backbone network,integrates the COT attention block into each residual block,and replaces all the 3×3 convolution kernels in the network,thereby constructing the residual network based on the COT attention.The COT module makes full use of the rich contextual information of adjacent keys to learn the attention matrix through the convolution operation of the multi-head attention mechanism,which enables the residual network to further focus on the discriminative regions of the image,and at the same time makes the overall network model easy to train.The experimental results on Deep Fashion and ACS datasets show that the method has fewer model parameters and computation,and higher recognition accuracy.