The Research Of Clothing Image Classification And Attribute Recognition Based On Deep Learning

With the development of society and the improvement of living standard,people’s pursuit of clothing is more and more diversified.The convenient online shopping has already become one of the main ways for people to choose clothing.The diversified demands require e-commerce platforms to label categories and attributes precisely for each picture,which can help the task of clothing item retrieval,clothing recommendation,etc.However,the cost of manual annotation is expensive because of the large number of images.With the development of computer vision technology,people begin to use the deep learning method to annotate categories and attributes of clothing images automatically.Because of the diversity of the clothing images,the current results of the classification and attribute recognition are not ideal.This paper uses DeepFashion,a large-scale dataset of clothing images with abundant annotation information,to study the classification and attribute recognition.This paper makes improvements based on an existing model of multi-task convolutional neural network.The main work of this paper can be summarized as follows:(1)Image preprocessing:The clothing target area is separated from the original image.To solve the imbalance of category distribution,we try to build a training dataset based on weighted random sampling technology,which should be helpful in training a more generalized classification model.(2)Estimate the key point locations of clothing:Following the original model,based on convolutional neural network,upsampling technology is used to restore the location information,which aims to predict the heat maps of clothing key point.This task is helpful to the tasks of image classification and attribute recognition.(3)Predict the category of clothing images:Following the original model,based on the predicted heat maps of clothing key point,attention mechanism is used to predict the category of clothing images.According to the characteristics of the dataset,this paper try to solve the problem of imbalanced category distribution and multi-class overfitting by adding focal loss function and label smoothing regularization,which improve the Top-3 accuracy rate of classification by 0.6 and 0.5 percentage points respectively.(4)Predict the attribute tags of clothing images:The method of attribute prediction is similar to the category prediction,but attributes are more diversified.Based on the original model,this paper makes improvements in two directions.First,multi-branch structure divided by attribute groups is built to learn the differential feature expression.Second,global average pooling,which can capture the global spatial information and actual meaning of the feature map,is used to replace fully connection layer.The results show that the multi-branch structure improves the Top-3 recall rate of attribute recognition by 1.7 percentage points,the global average pooling improves the Top-3 recall rate by 2.4 percentage points,and the combination of both improves the Top-3 recall rate by 3 percentage points.