Research On Clothing Image Retrieval Technology Based On Convolutional Neural Networks

In recent years,the Internet and e-commerce have developed at a high speed.People's shopping methods have gradually shifted from offline physical stores to online online stores,becoming mainstream shopping methods.Among,the online transaction volume of the electronic clothing industry is increasing year by year.In order to meet consumer demand for clothing,the e-commerce needs to continuously increase the types and quantity of clothing,which will cause the difficulty of the retrieval engine to accurately retrieve clothing products.At present,clothing image retrieval methods are mainly divided into two types:text-based methods and image content-based methods.Disadvantages of the text-based method:text descriptions between consumers and clothing merchants may be inconsistent,there are differences in individual understanding,and text information is difficult to describe the content of clothing images in detail.Disadvantages of image content-based methods:Traditional image content-based clothing image retrieval methods rely too much on the choice of image features,requiring researchers to have extensive industry experience.Therefore,it is of practical significance to explore an effective clothing image retrieval method.In view of the outstanding achievements of deep learning in the image field in recent years,this paper proposes a clothing image retrieval method based on convolutional neural networks.Under the premise of the current research status of clothing image retrieval,this article is investigated.First of all,a multi-scale SE-Xception network model is proposed to solve the problem that the richness of deep convolution feature information of clothing images is low,which leads to poor retrieval results.The experimental results show that the model is better than the commonly used CNN model in the clothing image classification and retrieval tasks.The retrieval subject framework borrows the deep binary hash fast retrieval method,and makes relevant adjustments and improvements based on the characteristics of clothing images:Use clothing semantic attributes to narrow the search area;Add triple loss to loss function;In order to further improve the retrieval accuracy,a deep convolution fusion feature that uses PCA dimensionality reduction technology is used.The main work and research results of this article include the following points:(1)A multi-scale depth separable convolution.Aiming at the problems of single-scale separable convolution and low feature information richness,a multi-scale separable convolution is proposed.First replace the 3×3 convolution kernel in the structure with 1×1,3×3,5×5 and Max Pooling.Before the multi-scale convolution structure,1×1 convolution is used to reduce the number of channels to reduce the amount of calculation.Multi-scale convolution experiments show that multi-scale separable convolution can effectively improve the accuracy of clothing image classification and retrieval(2)Multi-scale SE-Xception model structure.In order to further improve the classification effect of the model,a multi-scale separable convolution,Xception model and SE-Net model are organically combined.The experimental results show that the classification retrieval performance of the model in this paper is stronger than other models.(3)The overall retrieval framework combines the characteristics of clothing images and deep binary hash retrieval methods.At present,the number of clothing images is huge,and it is more convenient and effective to use deep learning models to automatically learn data distribution and extract features.Triplet Loss was added when training the model,which greatly improved the retrieval accuracy.The fusion of middle and deep features further improves the retrieval effect.(4)Comparison of clotthing image retrieval experiments of VGG network,residual network,Xception network,multi-scale Xception network and multi-scale SE-Xception network.The experimental results show that the multi-scale SE-Xception model of clothing image retrieval experiments are superior to the common CNN network.