Research On Clothing Image Retrieval By Region Of Interest And Feature Fusion

Online retail has become an indispensable part of residents' lives,and its sales volume has firstly exceeded 9 trillion yuan in 2018.The clothing product is one of the largest demand of the online retail,it is an important task for most e-commerce platforms to make users quickly and effectively retrieve the desired clothing.The major text-based clothing retrieval methods are easily affected by the subjective factors of image description.Therefore,a systematic study on content-based clothing image retrieval method is carried out,and a deep learning method based on interested regions and features fusion is proposed in this paper.In order to eliminate the influence of the background,we first extract the interested clothing regions.Considering the loss of clothing key points is easily caused by the conventional target detection algorithm,an reasoning algorithm is designed to obtain the interested clothing region with the higher inclusion ration of the key points.Then,the precise positioning of key points can be guaranteed in the interested clothing region,which is the preparation to extract local information around the key points.The low coordinate positioning accuracy of the clothing key points can be caused by the regression method.In this paper,the response value of the feature graph is adopted to calculate the position information of the key points,and a deconvolution module is added to the back-end of the network model to carry out up-sampling to improve the resolution of the feature graph,and further obtain higher positioning accuracy.Based on the above-mentioned two steps,we can extract the global features of the interested clothing region and the local features around each key point.All features are then fused with the constrained channel to retrieve clothing images.The performance with the constrained channel is superior to the features fusion without the constrained channel.Finally,considering the contradiction between the retrieval efficiency and accuracy in the clothing retrieval task,we propose a multi-task clothing retrieval method based on the weighted hash feature and clothing style classification.This method can make a flexible choice between retrieval efficiency and accuracy by specific requirements.The performance of the proposed algorithms is verified by experiments in the public datasets such as Inshop and Consumer2 shop.