Research And Implementation Of A Deep Learning-based Trend Prediction System For Fashion Elements In Clothing

The apparel industry is growing rapidly and occupies an important position in the global economy.Apparel companies need to quickly and accurately understand the current market trends in clothing elements in order to constantly adjust their production strategies to meet the market.Especially under the influence of fast fashion,this ability to know the trend of fashion elements in advance becomes more and more important.Based on the project of the Dongguan Humen Garment Collaborative Innovation Center’s Garment Cloud Design and Trading Platform,this paper designs and implements a garment trend prediction system based on deep learning technology to provide garment manufacturers and consumers with real-time access to and analysis of future garment trends.The problem with current research on apparel trend forecasting is that it can only model simple apparel trends,and most of the research has ignored the expertise in the apparel field.The trend prediction of fashion elements is itself a professional research direction in the field of clothing,so it is necessary to conduct the relevant research according to the characteristics of the trend data of fashion elements.Firstly,this paper constructs the coexistence relationship between clothing fashion elements based on the characteristic that certain clothing fashion elements will appear at the same time.Secondly,this paper proposes an attention mechanism based on the trend similarity to model the trend data of clothing fashion elements based on the characteristics that the trend of clothing fashion elements is cyclical and has a repetitive and similar life cycle.The main research of this paper is as follows:（1）Firstly,this paper proposes an Efficient Net model based on neural architecture search network as a clothing attribute predictor,which can automatically obtain the element attributes on clothing images quickly and accurately,and provide data support for the next trend prediction.Then,based on the differences in the views of different people in different regions on current clothing trends,this paper groups clothing fashion elements according to gender and region in user information and then performs statistics to predict the future direction of change of different clothing fashion element trends for each user group.The grouped trend data are used to build a model based on the Bi LSTM encoder-decoder framework,and the trend prediction is assisted by the coexistence relationship between fashion elements.The experimental results show that the proposed model can accurately predict future trends and outperforms existing methods.（2）The existence of redundant trend information in the long historical trend information will have an impact on the trend prediction results,and based on the cyclical nature of the trend changes of fashion elements,this paper proposes an attention mechanism based on trend similarity.By comparing the similarity of historical information with the trend of the recent period,the next trend information of similar historical information is used as an important influence on the future change trend.The prediction model based on the Bi LSTM encoder-decoder framework constructed by this method is able to predict the trend of fashion elements of clothing accurately.Compared with the latest KERN model to the half-year,the MAE and MAPE are improved by 5.09% and 4.54%,respectively.（3）Through the study of fashion element trend prediction,a fashion element trend prediction system is designed and implemented in this paper.It can automatically extract the attributes of fashion elements on clothing images and integrate the user information with the images to group the fashion element trends.Based on the historical data,the system can accurately predict the future trend change direction of fashion elements and provide users with queries according to different groups of people.