Female Fashion Style Analysis And Evaluation Based On Deep Learning

Fashion images contains various features,there are many researches on automatic classification and recognition based on fashion images.In contrast,there are few researches on the classification and evaluation of fashion style based on deep learning.In view of this,this paper uses different brands of fashion images as the research target to explore the reasonable representation of their visual style to classify the fashion style of different brands.Fashion style has its own unique design style.For example,the fashion style of Chanel is simple and exquisite,low-key and elegant;The fashion style of Dior is elegant,noble and gorgeous;The fashion style of Gucci is retro and classical.In previous researches,the definition of fashion style is usually by words,such as low-key,elegant,literary,retro,classical,etc.In the process of evaluation,professionals with rich experience are usually required to participate in the subjective evaluation.In order to improve the objectivity of evaluation,this paper proposes a method of automatic feature extraction,recognition and classification of brand fashion image by using convolution neural network,that is,the fashion style feature is transformed into the form of feature vector.The main content of this paper is divided into three parts: 1)the construction of fashion image dataset;2)the classification task based on fashion style solved by traditional image classification model;3)the classification task based on fashion style solved by few shot learning model.The specific research contents are as follows:1)We establish a dataset which contains 50 different kinds of brands fashion images.All the images in the dataset are from VOGUE.Among them,the dataset used in the traditional image classification method contains 50 categories of clothing images of different brands,600 pieces of each category.The training set,validation set and test set are randomly selected according to the ratio of 8:1:1.The data set used in few shot learning method contains 50 brands of clothing images,30 for each category.36 categories are randomly selected as training set,and the remaining 14 categories are used as support set and query set.2)We use ResNet-18、ResNet-34、ResNet-50、Efficient Net-B0、Efficient Net-B1、Efficient NetB2、Efficient Net-B3 to realize the classification task of different brands of fashion style.The experimental results show that convolutional neural network can recognize and classify different brands of clothing,and the classification accuracy of Efficientnet-b3 model is the highest,up to(0.9655).The results show that if the learning rate is set too small,the convergence process of the loss function is slow;if the learning rate is set too large,the gradient will oscillate back and forth near the minimum value,or even cannot converge.The weight attenuation coefficient and image size also have influence on the discrimination results,so the best parameters should be selected according to the experimental environment.3)We use Siamese network,Prototype network and Meta baseline network to test our data set.Compared and analyzed the classification results of each network model used,verified the classification accuracy and human accuracy of the model of few shot learning in the case of few images data.The results show that in the case of few images data,it can be classified by using few shot learning method.The classification accuracy of using meta baseline network in 5-way,1-shot task is as high as(0.9475),which is much higher than that of human classification(0.5680).