The Research On Multi-product Image Classification Based On Convolutional Neural Network

With the rapid development of The Internet of Things and e-commerce,online shopping has become an essential part of people's life.From the technical point of view,the effective retrieval and classification of product images based on the imaging features have been playing an extremely important role in support of e-commerce and related applications.The traditional classification method of products is based on text,which can not describe the overall characteristics of the goods,and manual labelling is time-consuming and laborious.Some content based classification methods are difficult to extract features,and the effect of classification is not good.Even though some automatic classification methods based on deep learning are put forward,it is also limited by the by the fact that the shallow network classification results are not good and the network training under small samples is not ideal.To solve these problems,this paper proposes two methods of products classification based on convolutional neural networks.The main work carried out is as follows:(1)Have a research on current classification technology of product images.Through reading a lot of literature about image classification,we analyzed the research background and current research situation of commodity image classification,and summarized various of classification methods used nowadays.(2)The basic principles of this study are introduced.Firstly,the development of neural network is introduced.Then the convolutional layer,the pooled layer,the fully connected layer,and the activation function of the convolutional neural network are analyzed.Finally,the common classification framework and feature extraction methods are listed.(3)Aiming at the poor classification effect of the shallow network,a product image classification method based on improved structure convolutional neural network is proposed.Compared with AlexNet,this structure optimizes the size of the convolution kernel,changes the connection of each layer,and reduces the connection parameters of the fully connected layer.The obtained numerical results show that the classification accuracy for the novel convolutional neural network structure is able to achieve up to 91.3%,which is even higher than that of 88.2% by the Alex Net.(4)Aiming at the problem of imperfect network training under small samples,a classification method of convolution neural network based on transfer learning is proposed.During data pre-processing,SVM algorithm is used to segment the background of the product image.In the training and fine-tuning process,a large dataset ImageNet2012 was used to pretrain the network structure.The pre-trained parameters were migrated to the new network as initialization parameters,and the training was completed with the pre-processed data.In the classification of 20 kinds of product images,the convolutional neural network classification method based on the pre-trained model achieves an accuracy of 97.6%.