Research On Garment Image Retrieval Technology Based On Deep Learning And Traditional Features

In recent years,with the development of the Internet and big data,the development of clothing e-commerce is more and more vigorous,the number of clothing images on the Internet is growing rapidly,more and more people choose online shopping to replace the physical store shopping,so the demand of users for clothing image retrieval is also growing.At present,there are two methods for garment image retrieval:text search and image search.The clothing image retrieval method based on text search map is becoming increasingly tired in the face of massive data due to the limited expression ability of key words and the limited clothing characteristics it contains.However,the current clothing retrieval method of image search relies on the semantic information of clothing images,and the annotation of these semantic information is subjective without a unified standard and contains a limited amount of information,which is unsatisfactory when faced with big data.Therefore,it is of great significance to improve the performance of clothing image classification retrieval method.Based on deep learning and traditional features,this paper studies the classification and retrieval technology of clothing images,mainly completing the following work:1.The key step of image search is to extract the feature vectors of clothing images.This paper first introduces the traditional representation of image features and the deep learning representation.In the traditional feature representation.HOG(Histogram of Oriented Gradient)feature,color histogram feature,edge histogram feature,Daisy feature and Gabor feature are mainly introduced.In the representation of deep learning,the theoretical basis of deep learning and common neural network structure are introduced.2.In addition to low-level information such as color,clothing image also has some special attribute information,such as clothing type,length,texture,sleeve type,collar type,etc.,and the clothing style is expressed by these attributes.In order to classify these special clothing attributes,this paper trained a deep convolutional neural network to classify them.When extracting clothing image features,a multi-label classification model of clothing attributes based on convolutional neural network was used.Due to the uneven distribution of clothing data with each special attribute in the clothing data set,this paper adopts the migration learning method to retrain the network model.Experimental results show that the trained network model can accurately extract and represent the attribute features,and obtain better attribute classification results.3.Clothing image is easy to be affected by factors such as background,illumination and deformation,etc.to extract the feature vector of clothing image.To solve this problem,this paper introduces similarity measurement learning,so as to optimize the clothing features extracted by neural network.The experimental results show that the retrieval effect is improved after the similarity learning optimization.4.The features extracted by deep learning are all deep semantic features and lack shallow features,which makes the retrieval effect not ideal enough.In this paper,re-ranking image retrieval method combined with traditional features is put forward.In this paper,features extracted by convolutional neural network and image features extracted by traditional methods are combined into re-ranking,so as to accurately locate,screen and sort the initial retrieval results.Experimental results show that this method can improve the accuracy of garment retrieval effectively.