Research On Clothing Compatibility Prediction Based On Graph Convolution Neural Network

With the improvement of people's living standards and the rapid development of the apparel retail,consumers' demand for fashion apparel recommendations is also increasing.The purpose of this topic is to answer a very practical question: "Which apparel should be chosen to match a given apparel and form a reasonable outfit?" The key to this problem is how to estimate the matching relationship between clothing.Judging the compatibility of outfits is a very challenging task.Determining whether a set of clothes fit together is not only an aesthetic judgment.Generally,it involves understanding the visual style of clothing,and even the social and cultural background.This task is also the basis for many other apparel industry applications,such as personalized fashion design,apparel creation and fashion trend prediction.However,the concept of fashion is elusive and depends on a variety of human subjective concepts.All attributes vary from person to person and will change over time.In the previous related work,most of them only considered the collocation between the two items,and failed to make full use of the complex relationship between each item of clothing in a outfit.The main research content of this article proposes improvements to the current clothing collocation prediction method.The first is to introduce the collocation relationship context information when predicting the collocation relationship between costumes.The relationship between the to-be-predicted costumes and other related costumes can be expressed as a graph structure and modeled using a graph convolutional neural network model.In further research,it is also considered that such rich context information is usually difficult to obtain,so in subsequent research,it is proposed to first extract the dependency relationships between different attributes of clothing based on other data,and guide the model by establishing a knowledge graph of these attributes.Good to get the feature embedding of a single clothing image.Furthermore,in order to consider the interdependence between costumes in the same set of collocations,the method of representing the combination as a sequence model in the previous study is improved,and the graph attention model is used for modeling,and the score for the entire collocation is finally obtained.And based on these studies,a set of clothing matchingrecommendation system was designed and implemented.This article also conducts experiments on the following two tasks for the proposed model:(1)Fill in the blanks: choose the clothing that matches the existing clothing combination;(2)Compatibility prediction: predict the reasonable degree of collocation for a given clothing.The experimental results prove that our proposed method performs better than other existing methods on multiple data sets.