Reasearch On The Clothing Matching Method Based On MBTI Personality Theory

In modern society,clothing accessories are more and more endowed with social significance and respect connotation.It has become a consensus of many people to choose reasonable clothing on specific occasions.Therefore,in order to help people complete the complex issue of clothing matching,scholars put forward various theories and methods.On the one hand,professionals in the accumulation of practical experience summed up the collocation rules,which express the image characteristics of the wearer’s heart by clothing as the media,so the clothing also presents the inner characteristics of the collocation to a certain extent.At the same time,some scholars have studied the relationship between clothing factors and the personality characteristics expressed,but there are some problems such as limited to single factor,lack of overall perspective,and incomplete description dimension of personality model.On the other hand,scholars at home and abroad pay more and more attention to the research progress of the whole style of clothing.At present,most of the algorithms are based on the clothing itself,through image and text description for machine learning training,to a certain extent,it ignores the emotional semantics represented by clothing features,and the matching objects are mostly pairs of single items,lacking the matching algorithm research of complete sets of clothing including accessories.In view of the above problems,this paper establishes a clothing matching method based on the Myers Briggs Type Index(MBTI)personality type,which will interpret the personality characteristics of clothing collocation from the perspective of others,and help consumers choose clothes that can convey their expected personality image.Firstly,the subjective evaluation experiment of MBTI personality type was carried out,and the evaluation personnel with certain representative and aesthetic level were invited to select MBTI personality type evaluation value for each suit of clothing.At the same time,the image features of clothing items are extracted by the fine-tuning resnet50,and the text features are extracted by word bag model and principal component analysis,and then the feature vector of clothing set is obtained by fusion.Taking the feature vector of the suit as the input and the corresponding subjective evaluation value of MBTI personality type as the output,the neural network was constructed and trained to compare the prediction accuracy of personality type from different modes,different categories and different personality dimensions.Finally,the prediction model is applied to specific clothing matching tasks,including matching degree measurement and clothing recommendation retrieval,and the results are compared with the previous Euclidean distance model and point multiplication model.The main conclusions are as follows:(1)The MBTI personality type subjective evaluation experiment was carried out to verify the validity of the data.The data of subjective evaluation experiment were Polyvore outputs published by former people 2046 sets of clothing were randomly selected from the 2018 clothing data set,with a total of 10620 pieces,which were divided into eight categories.At the same time,an evaluation platform was built.According to the questionnaire results,five people with high fashion taste,rich experience in collocation and different professional backgrounds were selected to participate in the experiment.Finally,the difference and consistency analysis of the evaluation results showed that the highest consistency score was 0.753 and the result was significant The evaluation data can be used for the subsequent neural network training.(2)The image and text features of clothing items are extracted and fused into suit features.Resnet50 with fine-tuning accuracy of 0.82 is used as the model of image feature extraction.After inputting the image of single item,2048 dimensional vector before FC is extracted as image feature of single item.At the same time,word bag model and principal component analysis(PCA)are used for word segmentation,filtering,vectorization,standardization and dimensionality reduction.2048 dimensional text features are successfully extracted.Finally,feature fusion step is used with the addition method In this method,the clothing features of the same category are added directly,and the clothing features of different categories are merged by concatenate.The features of different modes are stored in different channels,and finally a set of clothing feature vectors is obtained.(3)The MBTI personality type prediction model of clothing was established and the influence of different factors on the model was analyzed.The goal of network training is to regress the input set image and text feature vector into personality type four-dimensional data.After 30 times of training(epoch),the loss function reaches convergence,and the accuracy rate of verification set increases from 0.5 to 0.75.The model training effect is remarkable.Comparative experiments are carried out and charts are obtained by combining traditional statistical methods to explore different modes(image,text)and different products The results show that the image information extracted by deep neural network is more helpful to establish the relationship between clothing and personality characteristics,among which the three elements of color are related to the tendency of personality type;the influence degree of different categories on personality type is shoes > clothing > Accessories > bag,and the secondary category is related to people Among different personality types,the "feeling & intuition" dimension is easier to predict due to the data concentration,and the "extraversion & Introversion" distribution is even,and the prediction is more difficult.Therefore,the distribution of data itself will also affect the accuracy of prediction.(4)Based on MBTI personality type prediction model,the specific task of clothing matching is realized and compared with previous models.The prediction model is applied to the specific clothing matching task,and compared with the previous models.The results show that in the matching degree measurement task,the matching degree output of personality model on some clothing conforms to the basic facts,the output of distance model and point multiplication model is quite different from subjective intuition,and the quantitative effect of collocation degree is poor.In the whole data set,the AUC value of personality model is 0.6,51 and 0.49 of the other two models,the performance was better;in the retrieval and recommendation task,the recommendation ranking of personality model output on some clothing was more accurate,and the recommendation effect of distance model and point multiplication model was poor.In the whole data set,when the number of candidate items was large,the retrieval recommendation performance index MRR of personality model was 0.11,which was better than that of distance model and point multiplication model 0.02.