| In recent years,with the booming development of e-commerce,the number of clothing images has been on a steep rise.It has become increasingly difficult for users to retrieve their favourite clothing from the huge amount of clothing images,so how to find out exactly what consumers want in tens of thousands of clothing images has become an urgent need.Text-based apparel image retrieval requires workforce to annotate.It is subjective,and the text annotation can hardly completely summarise all the information in the image,resulting in unsatisfactory retrieval results.The currently emerging "image search" garment retrieval technology relies on image content features for retrieval,although it avoids the trouble of manual labeling,it is difficult to highlight detailed features such as patterns and patterns in local areas of garments,and there is also the problem of partial loss of image information,resulting in a low retrieval accuracy rate.This paper depends on the background of the group's clothing retrieval system and uses deep learning methods to achieve accurate retrieval of clothing images.The main work of this paper is as follows.(1)A hybrid attention mechanism-based feature extraction network for clothing images is proposed.Aiming at the problem that clothing images have a large amount of semantic information and detail information,and the features extracted by most convolutional networks are the overall features of clothing images and lack the features that highlight the local characteristics of clothes,this paper proposes a feature extraction network CSP-Net based on a hybrid attention mechanism,which enhances the attention to local detail features such as patterns and patterns in essential regions of clothes by introducing a hybrid attention mechanism to strengthen feature representation.(2)The loss function of the CSP-Net is optimised.The addition of ternary group loss to the cross-entropy loss function only enables the model to enhance the training of clothing samples with small differences,strengthen the recognition ability of small difference images,and improve the recognition of different clothing image features.The experimental results show that the check accuracy and recall rate of the network with the addition of ternary group loss are 1.9% and 2.2%higher than those of the network using only the cross-entropy loss function.(3)A clothing image retrieval model based on multilayer feature fusion is proposed.CSPNet enhances the attention to detail features such as patterns and patterns in local regions of clothes based on the overall features obtained from clothing images,but also weakens the image detail information such as color and texture in the bottom layer and style details in the middle layer extracted by the network,and there is a problem of partial image information loss.Therefore,we further propose a multi-layer feature fusion-based clothing image retrieval model MF-CSPNet,which extracts the image detail information of the bottom and middle layers of the CSP-Net network with the semantic information of the top layer,and uses the feature fusion by multi-scale convolutional kernel method to extract features of different scale sizes while fusing multi-layer features,so that the extracted image features are more comprehensive and realize the clothing image accurate retrieval.The experimental results show that the proposed apparel image retrieval model has 2.4%,2.7%,and 2.8% higher search accuracy,recall,and top-5 accuracy than the CSP-Net model without feature fusion,respectively.(4)Finally,a clothing image retrieval system is constructed by combining the clothing image retrieval algorithm proposed in this paper,which can effectively handle a large number of clothing images and improve the intelligence of clothing image retrieval. |
PDF Full Text Request