Product Image Classification Based On Convolutional Sparse Representation And Convolutional Neural Network

With the vigorous development of e-commerce,online shopping has become a popular consumption behavior.As the main carrier of product information transmission,product image has become an important medium of selective purchase.The number of product image has a sharp increase in all kinds of shopping websites,which increases the difficulty for the user to pick the goods in massive products.Therefore,it is an urgent requirement for users and e-commerce development to provide an efficient product image classification method.In recent years,convolutional sparse representation and CNN(Convolutional Neural Network)has been widely applied in the research of image classification,which has become the research hotspot in the field of image processing and computer vision.According to the characteristics of product images,this paper studies product image classification based on convolutional sparse representation and CNN.The work is as follows:1.The consistency between image patches is ignored in the product image classification method based on traditional sparse representation.In order to solve this problem,a product image classification method based on CSC(Convolutional Sparse Coding)is proposed.Firstly,this method can directly input 2D images to train convolutional dictionary,and obtain the optimal sparse feature maps,which takes full account of the similarity betwwen image patches.Then,the final features of the product image are obtained by using max-pooling method which can reduce the dimension of the sparse feature maps.Finally,the SVM(Support Vector Machine)is used to accomplish the product image classification task.The experimental results show that this method improves the classification accuracy.2.CNN has obvious effect on the classification of product images,but it needs a lot of label data.In this paper,a product image classification method based on unsupervised CNN is adopted.In this method,the CNN model pre-trained by CSC algorithm can realize the unsupervised feature learning of product images.The layer-skipping strategy in the network ensures that the final features are consist of both high-level global and low-level local features,which improves the discriminability and reliability of the features.After the final features are obtained,the softmax regression is utilized to classify product types.The experimental results show that the CNN with layer-skipping strategy has some advantages in product image classification.