Research On Classification And Retrieval Of Clothing Image Based On Convolutional Neural Network

With the vigorous development of apparel e-commerce and the increase in online shopping transaction volume year by year,the amount of clothing image data on the internet has increased dramatically.How to rapid and effective automatic classification and retrieve of massive clothing images is an urgent problem for stakeholders in the apparel e-commerce industry to solve.This paper studies the classification and retrieval of clothing images based on convolutional neural network.We use feature fusion to achieve accurate classification and efficient retrieval of clothing images,which has extremely high industry commercial value and academic significance,and the main works are as follows:(1)A clothing image classification network based on improved VGG16 is proposed.By improving the VGG16 network structure and using multi-scale convolution for feature extraction,the richness of clothing image feature information is improved.Additionally,we use L2 regularization to enhance the anti-interference ability of the network.Experimental results show that the classification model based on the improved VGG16 has higher classification accuracy than other traditional classification methods such as VGG16.(2)In this paper,we propose a novel model,named Multi?XMNet,to solve the clothing images classification problem,which is composed of two CNN branches networks.Firstly,the single-scale depth separable convolution in Xception is replaced with 1x1,3x3,and 5x5 convolution kernels of different scales,and a multi-scale feature extraction network named Multi?X is obtained as a branch network of the model.Then the other extracts attention mechanism features from the whole expressional image by MobileNetV3-small network.Both multiscale and attention mechanism features are aggregated before making classification.Additionally,in the training stage,global average pooling and convolutional layers are used instead of the fully connected layer to classify the final features,which speed up model training and alleviates the problem of overfitting caused by too many parameters.The model is applied to the Deep Fashion dataset for experimental testing.The experimental results show that the classification accuracy of this model is 95.38%,which is better than InceptionV3,Xception,InceptionV3?Xception,MobileNetV3 and F-ResNet by 2.22%,5.58%,3.32%,2.27%,and 6.81%.(3)On the basis of clothing images classification based on feature fusion,this paper proposes a clothing image retrieval method that combines color features and depth features and implements a system prototype.The proposed method mainly consists of two parts.One branch extracts depth features of clothing images by Multi?XMNet,while the other extracts color histogram features from the whole expressional image.We aggregate the depth features and color histogram features,and use similarity measurement algorithms to retrieve highly similar clothing in the clothing database.Finally,we built a prototype of clothing image retrieval system based on the retrieval method proposed in this paper.The clothing image classification and retrieval model proposed in this paper can be used to help industry managers,researchers and consumers perform rapid and effective automatic classification and retrieval of clothing images.Additionally,this model can also help to establish image classification and retrieval models and systems in other scenarios.