Clothing Image Retrieval And Collocation Based On Deep Learning

With the development of the clothing e-commerce and the coming of the era of big data,vast amounts of clothing data are being stored on the Internet,the users' needs of clothing retrieval and collocation are increasing day by day.Helping users find their favorite clothing quickly and accurately,and recommend appropriate collocation to them are important measures adopted by clothing e-commerce shops to enhance user experience and sales.Therefore,study on clothing collocation technology and clothing collocation technology already becomes a meaningful task.Clothing retrieval and collocation all depend on the understanding of clothing image content information,which can be converted to the representation of clothing features.This thesis use deep learning method to represent the image features,and study their application of clothing retrieval and collocation.In this thesis,the main work is as follows:1.Two feature representation methods of clothing image are introduced in this thesis,including local features and deep learning.We analyze the characteristic of these features to provides a theoretical basis for clothing retrieval and collocation.For local features,we focus on the introduction of SIFT,SURF and BoF feature model which based on the first two features.For deep learning,we focus on the introduction of autoencoder and convolution neural network.2.A novel clothing attribute multi-label classification model based on convolution neural network is presented to extract features of clothing images.Clothing images are rich in clothing-specific attribute information,such as color,pattern,length of sleeves.We classify these clothing attributes by training a deep convolution neural network,and use the deep activation values of the neural network to represent the clothing features.In the process of training,in order to compensate for the lack of clothing training data,we use migration learning method to re-train the network.The experimental results show that the attributes of the clothing can be well represented by the features extracted by the deep convolution network,and a better clothing retrieval result is obtained.3.A convolution neural network based on similarity measure learning is presented to optimize the clothing features.The mothod includes Triplet similarity measure learning and a double convolution neural network,which using the triple clothing images to train network parameters,and to optimize the matching of clothing features.Then we use this network to extract clothing features,using approximate nearest neighbor algorithm to find similar clothing.The experimental results show that the optimized features improve the retrival results in the commodity clothing images and the photos take by users,having high robust to illumination and deformation.4.A novel clothing collocation feature based on deep learning feature and local feature is presented to build collocation space.Firstly,we use convolution neural network and BoF model to extract features,and then we use denoising autoencoder to reduce the dimension of clothing features.Lastly,we construct association rules based on frequent itemsets,and build collocation space using data collocation set.The experimental results show that the collocation space can be used to find the suitable clothing collocations for users.