Image Classification Method Based On CNN And Its Application In E-Commerce Product Image Classification

With the advancement and development of the Internet,people are increasingly inclined to purchase goods through online shopping.In the past few years,the variety of commodities has increased dramatically.How to make consumers be able to find related commodities quickly in mass commodities is an urgent problem to be solved.Although the traditional classification method based on text keywords is convenient and fast,it is prone to misclassification due to the one-sidedness of text annotation information.The product image contains a wealth of information and data that can visually reveal most of the features of the product.The product image classification method based on convolutional neural network can provide customers and merchants with a better product inquiry and retrieval experience,and help e-commerce platform to recommend products.The main research contents of this thesis are as follows:(1)The derivative of the commonly used activation function relu is always zero at the X negative axis,which leads to the “necrosis” of neurons in the training process.And the existing combined activation function relu-softplus has a low convergence rate due to the low learning rate in the case of model convergence.In view of the above problems,we design and implement a new combinatorial activation function relusoftsign.The role of activation function in training process is analyzed in detail,and the key points that need to be considered in the design of activation function are given.Then according to these points,the relu and softsign functions are combined in the positive and negative half axes of X axis,so that the derivative of X and negative half axes is no longer constant to zero.The experimental results on the MNIST,PI100,CIFAR-100 and Caltech256 datasets show that the use of the relu-softsign combined activation function improves the model classification accuracy,simply and effectively alleviates the irreversible “necrosis” of neurons.At the same time,the model convergence speed is accelerated,and the convergence performance of the combined function is better on complex data sets.(2)Dropout is a convenient and effective method to prevent overfitting,the drop probability set for the method works for all neurons in the layer,which causes some useful information to be discarded,resulting in the decrease of the average training accuracy of the model.In response to this problem,the Sep-Dropout method is proposed.First,the neurons are divided into two matrices based on their importance,and then the important neuron matrix is zeroed with a lower probability,the unimportant neuron matrix is zeroed with a higher probability.Finally,the two matrices are integrated.Therefore,the method can effectively reduce the possibility that important neurons are discarded,and improve classification accuracy.Experimental results on the MNIST,PI100,CIFAR-100 and Caltech256 datasets show that the SepDropout method can prevent over-fitting,and have higher classification accuracy than the Dropout method at the same time.(3)The relu-softsign function and the Sep-Dropout method proposed in this thesis are applied to E-commerce image datasets for comprehensive experiments.First,analyze the characteristics of authoritative datasets in the field of e-commerce image classification,build common product self-built datasets and fine-ware product selfbuilt datasets based on its structure,and then preprocess the image datasets.Experimental results on the PI100 dataset and the two self-built datasets show that on the E-commerce image datesets,the relu-softsign function can improve the classification accuracy of the model on both the training set and the test set,the SepDropout method can guarantee the accuracy of the train set while preventing overfitting.And the model applied the two methods achieves higher classification accuracy in both the training set and the test set.Finally,the final model applied the relu-softsign function and the Sep-Dropout method is compared with other researchers' E-commerce image classification methods on the PI100 dataset.The experimental results show that the method of this thesis achieves higher accuracy and better classification result.