Clothing Image Classification And Retrieval Based On Deep Learning

With the development of apparel electronic,the amount of clothing image data on the Internet has increased dramatically.Facing massive clothing image data,it would cost a lot of manpower and time to label semantic attributes of garment images manually.Moreover,semantic attributes can not fully express the rich information in clothing images which would result in poor retrieval result.Therefore,it is a meaningful topic to find an efficient and effective method for clothing image classification and retrieval.Recently,deep learning has made outstanding achievements in image processing.In this paper,we study the classification and retrieval tehnology of clothing image based on deep learning,and the main works are as follows:Firstly,we briefly introduce the basic concepts of deep learning and the structures of frequently-used models,and expounds the principles and model components of convolutional neural network in detail.To meet the requirement of classifying multiple attributes of clothing image simultaneously,a convolutional neural network based on multi-task learning is proposed.In order to overcome the influence of background,lighting,deformation and other factors,a convolution neural network with metric learning is proposed,namely the Siamese network and Triplet network.Experimental results show that the introduction of metric learning,especially the Triplet network,can significantly improve the accuracy of classification.Then,We use the trained convolution neural network for feature extraction,and compare with traditional methods based on SIFT features.Experimental results show that the extracted features using convolutional neural networks have obvious advantages over SIFT features.Especially,the anti-interference ability of convolutional neural network based on Triplet structure is improved obviously.In addition,the fusion of feature maps from both convolution layer and fully connected layer has better effect.To eliminate the impact of the background even further,we also use Faster-RCNN for clothing detection,and extract of the region of interest.At last,two simple methods are proposed to improve retrieval speed,one is K-means clustering,and the other one is pre-classification with semantic attributes prediction.And experimental results show that K-means clustering is the better choice.