Research On Clothing Style Classification Method Based On Deep Learning

E-commerce platforms are gaining momentum,and the vast amount of product image data in e-commerce platforms contains a wealth of valuable information.In the case of clothing products,consumers are not only interested in information on practical attributes,but also on style attributes.It is therefore relevant to use images of clothing products as data for identification and mining.Currently there is a wealth of research on the classification of practical attributes of clothing,but not much research has been done on the classification of clothing style characteristics.Moreover,feature engineering is not enough to solve the problem of style recognition.Based on the above analysis,this paper takes clothing image as the basic data to explore the quantitative method of style features.While using traditional feature engineering,both shallow features and deep features are taken into consideration,and then the classification task based on the fusion of basic features and deep features is completed.As consumers pay more and more attention to the pursuit of personalized wear,clothing style attributes and practical attributes have the same important status.At present,there is little research on the task of clothing style classification,and there is a lack of public and reliable data set auxiliary method research;There is little research on the characteristics of clothing style,and the methods for style classification need to be explored urgently;The style defining elements include objective defining elements and subjective defining elements.The objective defining elements include shape,color and other characteristics,but it is difficult to represent the style characteristics and thus achieve image style classification.How to combine the subjective and objective defining elements to jointly achieve the evaluation of style characteristics and help to improve the accuracy of style recognition remains to be studied.In order to solve the above problems,the main contents and results of this study are as follows:(1)Construct clothing image style data set.According to the existing literature and relevant knowledge,the influencing factors of clothing style classification are summarized,and the influencing factors of style characteristics are obtained.The images in the DeepFashion clothing image public dataset are re-labeled into 8 categories to create the clothing style dataset,which lays the foundation for the implementation of the subsequent clothing image style classification model.(2)The realization of quantification of clothing style characteristics.Basic visual features,shallow features and depth features are extracted.According to different feature combinations,Softmax classifier is used to classify and determine the style feature composition suitable for style classification.In addition,compare the style recognition ability of Alex net,Inception V3,ResNet50 and VGG16 networks on the data set constructed in this paper.Filter out extractors suitable for extracting style features.(3)Construct a clothing style classification model based on VGG16 extractor.Feature engineering extracts basic visual features,i.e.colour,shape and texture basic features by HSV colour space,Fourier description method and circular LBP respectively.And the basic features are strengthened by secondary feature extraction through shallow neural network,so that the style-related features are extracted automatically.VGG16 extracts the depth features of the image and fuses them with the shallow features to obtain a suitable convolutional neural network style classification model.The output shallow style features are extracted twice and fused with the depth features obtained from deep learning and put into the Softmax classifier to achieve the classification task.The model in this paper can achieve a classification result of 72.26%.Finally,the style classification ability of the unsupervised learning method based on DeepCluster network is compared with the supervised learning method in this paper to verify the effectiveness of the model in this paper.