Narrow Pooling Clothing Classification Based On Attention Mechanism

With the rapid popularity of e-commerce and online shopping in the clothing industry,accurate and efficient clothing image classification can not only realize automatic labeling,but also greatly improve the efficiency of clothing applications such as clothing retrieval and virtual fitting.Due to the complex and large number of clothing image scenes,and susceptible to factors such as lighting,human posture and deformation,traditional apparel classification algorithms such as support vector machines and K-nearest neighbor algorithms are not efficient and have serious misclassification problems.Therefore,how to use deep learning methods to achieve accurate classification of clothing has become an important element in the current image classification research field.In this paper,we propose an improved classification algorithm based on asymmetric convolution,hybrid attention mechanism and bar pooling for clothing image classification based on the problems that classification models such as VGG16 and Res Net of clothing image will lose the color and shape of the shallow layer in the deep network,and the feature information is not fully utilized and will be disturbed by the background information in the feature learning process.The main contents of the research work include:(1)The DLA-VGG16 classification model is proposed to address the problem that deep convolutional clothing classification networks lose more shallow features when high-level feature information is obtained.The model is based on iterative depth aggregation and hierarchical depth aggregation to improve the classification model of VGG16 base network,by adding depth aggregation points of the intra-block convolutional layers in each convolutional block and adding aggregation modules between blocks to make the classification model make full use of different sizes of perceptual fields to obtain deep features of clothing while reducing the loss of shallow feature information,thus reducing the loss of clothing feature information in the learning process of the network due to up-sampling and down-sampling in the learning process of the network.(2)In response to the fact that deep convolutional networks using attention mechanism for clothing classification usually focus only on two-dimensional feature information and ignore the interactive utilization of three-dimensional feature information,a hybrid attention module based on global three-dimensional information is proposed.This module adds the hybrid attention module to the first and fourth layers of the Rea Net network for global information interaction,thus enhancing the utilization of channel weights of cross-dimensional information.By expanding the channel-space dependence of cross-dimensional feature information,the weights of feature information can be better assigned,and thus feature information with higher correlation with the results of garment image classification tasks can be effectively extracted.(3)A Res Net-based rectangular pooling layer classification method is proposed to address the problem that clothing classification is susceptible to interference by background information and the misclassification of clothing categories with high similarity.The method uses a rectangular pooling layer instead of the traditional square pooling layer in the network to obtain the global information of clothing images by scanning the global along the longer dimensional direction,thus effectively avoiding the loss caused by the square pooling layer in extracting features.At the same time,local features are further extracted by narrower pooling layers in the block of Res Net,which reduces the extraction of background features unrelated to clothing classification.By validating the related algorithms on the Fashion-Mnist and Deep Fashion clothing datasets,the asymmetric convolution-based deep aggregated clothing image classification method proposed in the paper shows an accuracy improvement of 1.22% and 4.12%,respectively,compared to the VGG16 network,and the improved rectangular pooled Res Net clothing classification based on the hybrid attention mechanism method improves 0.55% and 1.14%,respectively,compared to the base network.The comprehensive experimental results show that the proposed attention mechanism-based striped pooled clothing classification algorithm can better fuse global and local,deep and shallow features of clothing images,while combining the hybrid attention module to improve the classification accuracy of clothing images.