Research On Clothing Image Semantic Retrieval Based On Residual Network

In recent years,with the rapid development of E-commerce,clothing image data has shown explosive growth.In massive clothing images,it is a subject of great commercial value and academic significance to retrieve the target clothing efficiently and accurately.The traditional clothing retrieval method has low efficiency and large error,which is not conducive to popularization.With the development of artificial intelligence,clothing image retrieval based on deep learning has become a hot research topic.This paper focuses on the feature extraction method and distance measurement algorithm based on residual network.The main research contents of this paper are as follows（1）A new Clothing feature extraction model based on Dilated Convolutional Residual Networks（DCRN）is proposed.DCRN model integrates dilated convolution into residual network,so as to improve the extraction ability of deep semantic information of the network.Firstly,DCRN model uses STEM module to extract shallow detail information;Secondly,the dilated convolution and residual network are fused through DCRN module;Finally,the highdimensional feature vector is obtained through the binary vector module.Experiments show that the fusion dilated convolution improves the extraction ability of the residual network for deep semantic features,reduces the number of parameters and improves the computational efficiency.（2）A clothing feature extraction model（Transformer-Dilated Convolutional Residual Networks,T-DCRN）integrating transformer encoder and DCRN network is proposed.The method of this paper is to integrate the Transformer-Encoder module into the DCRN network.Firstly,the feature map is encoded;Secondly,clothing image features are extracted through six groups of multi head attention mechanism layer and feedforward network.Experiments show that compared with other residual networks,due to the integration of transformer encoder,TDCRN network has enhanced the perception ability of clothing component relative position information,so as to improve the ability of deep semantic feature extraction,which reflects the advantages of T-DCRN network in clothing style understanding.（3）In the framework of K-means algorithm,a Mixed Distance measurement algorithm（MD）is proposed.This paper integrates Markov distance and cosine distance,and analyzes their combination.Experiments show that the spatial distance of feature vector can be calculated stably and efficiently by calculating the sum of cosine distance and Mahalanobis distance;The super parameters in the sum of the two are analyzed λ Impact on search results.In order to test the performance of the model and algorithm proposed in this paper,a number of experiments are carried out on the Deep Fashion data set to evaluate the clothing image retrieval method proposed in this paper.The experimental results show that the model and algorithm can better solve the problem of clothing style recognition and obtain better retrieval accuracy.This paper hopes to improve the efficiency of clothing retrieval by using the latest artificial intelligence technology.