Research On Key Technologies Of Clothing Matching Recommendation Based On Deep Learning

As a symbol of soft power such as culture and creative design,the fashion industry contains significant market value and development potential.With the development of society,people’s fashion tastes and needs are constantly improving,and clothing matching is the most direct way to pursue fashion.Clothing matching is the combination product of knowledge in the fashion field and personal subjective creativity.Not everyone has the skills and experiences of professional clothing matching designers.This paper uses deep learning technology to learn high-quality clothing knowledge based on the large amount of professional clothing matching data from the professional clothing matching designer on the Internet,so as to realize intelligent clothing matching recommendation.For individual consumers,smart clothing matching recommendation can help save time and energy spent on dressing,and help improve dressing taste;for clothing e-commerce platforms,smart clothing matching recommendation can improve consumers’ online shopping experience,so as to enhance competitiveness;for fashion designers and production suppliers,through the analysis of the compatibility of people’s clothes in different regions and times,it can help to sort out the fashion factors that affect the clothing matching in different time and space,so as to provide references to improve the design,production and sales of more popular clothing products.Therefore,the research on the key technologies of clothing matching recommendation based on deep learning technology is of great significance.According to the clothing items provided by the user,the clothing matching recommendation recommends other items that complement the functions of the existing items and are visually compatible.Functional complementarity and visual compatibility are the basic elements for recommendation.This paper conducts research from two aspects: clothing type identification and clothing compatibility evaluation.Based on the Restnet50 residual structure,we design a feature fusion clothing type identification network,which fuses the shallow features and high-level semantic features extracted by multiple convolutional layers,and generate classification features through global average pooling,so as to realize the clothing types.The category recognition accuracy improves by 2.3 percentage points over the original Res Net50 classification network.For the evaluation of clothing compatibility,this paper proposes a clothing compatibility evaluation network FCET based on Transformer.FCET uses the segmentation of a single product image as the smallest feature input unit,so as to achieve a more fine-grained clothing matching compatibility evaluation;in addition,FCET uses a multi-head self-attention mechanism to capture the correlation between feature vectors,so as to mine the impact of different items on the overall compatibility,and improve the accuracy of the compatibility evaluation.In the FITB task,the accuracy of FCET is 4.2 percentage points higher than that of the LSTM-based compatibility evaluation network.Based on the FCET encoder output vector sequence and the overall fashion factor,we realize the compatibility diagnosis of dressing by calculating the influence weights of different image blocks and single items on the overall compatibility,thereby enhance the persuasiveness of the compatibility evaluation result.Based on the proposed clothing type identification technology and compatibility evaluation technology,this paper implements the clothing matching recommendation prototype system named Deep Rec,and displays the recommendationrelated functions through this system.Deep Rec can be applied to fashion APPs or clothing e-commerce platforms,so as to provide outfit suggestions for individual users while further enhance the competitiveness of online sales platforms.