Clothing Image Retrieval And Recommendation System Based On Perception And Reasoning Model

AI-based clothing e-commerce has become a hot research area due to the energetic development of online apparel market.Using machine learning methods to classify,retrieve and recommend clothing could promote the sales volume of online apparel shopping.However,existing approaches still have many shortcomings to be improved,such as robustness,interactivity,interpretability,etc.This article focuses on content-based clothing image retrieval and clothing recommendation based on high-level semantics.First,we propose a novel retrieval method to meet users' multi-dimensional requirements.It enables users to refine customized search conditions by modifying the color,texture,shape and attribute descriptors of the query images,thereby improving retrieval efficiency and quality.Compared with other contentbased retrieval methods,our approach extracts multiple types of low-level clothing descriptors from the clothing images and uses the innovate Hybrid Topic(HT)model to learn the high-level semantic representation of the above descriptors.Our model not only provides effective clothing retrieval features,but also could perform automatic image annotation by probabilistic reasoning.Second,we integrate both perception and reasoning models into a novel clothing recommendation system such that our recommended collocations simultaneously meet with human visual aesthetics,user's collocation experience and fashion sensitivity.To further improve the accuracy and robustness of our recommendation,we make full use of the multi-dimensional clothing information: including user purchase histories,clothing product images,merchant descriptions,street shot images from fashion websites and expert advices.When analyzing the above information,we use deep convolutional neural networks to extract visual and attribute descriptors from the clothing images.Meanwhile,we integrate and encapsulate these descriptors using our proposed hierarchical collocation model(HCM).In this way,we can learn the concept of style topics from apparel images,and explain the collocation pattern from a higher level of semantic knowledge.Last but not least,HCM structurally facilitates us to embed the expert knowledge of fashion authorities into the recommendation model,making our suggested collocations in line with the forefront of fashion.We conduct several experiments to evaluate the proposed clothing retrieval and recommendation approaches.The results show that our approaches achieve better performance than other state-of-the-art methods.In the clothing retrieval task,we designed an interactive search system based on the HT model —— "Magic Wardrobe".It supports user-defined query criteria and can evaluate the accuracy of HT algorithm under different experimental conditions.In the clothing recommendation task,we verified that the HCM algorithm can provide users with personalized and fashion-sensitive recommendation services while ensuring performance.Finally,many retrieval and recommendation examples are given to further confirm our conclusions.