Multi-Attribute Fusion Based Outfit Composition Recommendation

With the flourishing development of Internet technology,more and more consumers tend to shopping on e-commerce websites.According to recent research,a large proportion of the online orders are related to clothing.E-commerce websites does facilitates people's shopping activities to some extent,however,due to the abundance of clothing on e-commerce website,selecting clothing for outfit composition manually will not only increase consumers' shopping time,but also make online shopping tedious.In order to improve users,shopping efficiency and shopping experience,this paper proposes a clothing recommendation algorithm based on multi-attributes fusion.In this paper,visual information of clothing's color,texture and shape,as well as the high-order features derived from their fusion,are used as feature descriptors of clothing to complete the recommendation of fashion outfit composition.First of all,this paper carries out a series of image processing on clothing image,uses AutoEncoder to extract the superficial visual information of clothes,then maps the visual information to the interpretable semantic space and performs feature combinations using superficial visual information on semantic space,obtaining some combination features of clothes.Because these high-order features may contain some redundant or even negative feature information,this paper uses the heuristic search algorithm to filter the features that have a negative impact on the result by using the multi-branch structure of the Siamese Neural Network,thus obtaining the final feature descriptor of clothing image.Finally,a noise-insensitive algorithm is designed in this paper,which takes Euclidean distance between the descriptors of clothing as a measurement of the fashion outfit composition,so as to complete the recommendation task.In this paper,we use crawlers to crawl a great deal of clothing images,including tops,bottoms,shoes and hats,etc.from the Internet,and use the likes of outfit to represent the matching degree of clothing to construct a training set and a test set,and use these data to train the classifier.The experimental result shows that the addition of high-order features can improve the accuracy of fashion outfit composition analysis,while the feature selecting algorithm can further improve the final result.