Product Image Classification Based On Deep Learning

With the popularity of the Internet and varieties of terminal equipments, e-commerce has become a part of people’s lives, and it has been people’s urgent needs to achieve efficient image retrieval method. Deep learning can extract feature automatically based on image, and it can get deeper abstract features through multi-layers network samples the image. This thesis studies product image classification based on deep learning, and focuses on fine-grained classification of product image, the main work of this thesis is as follows:This thesis studies fine-grained classification of product image based on stack autoencode network(SAE), deep belief network(DBN), convolutional neural network(CNN)respectively. As traditional deep learning is time-consuming and requires higher hardware,product image fine-grained classification based on SAE, DBN, CNN combined with support vector machine(SVM) is proposed and implemented in this thesis. The results of experiment show the classification accuracies which SAE and DBN combined with SVM have achieved are a little better than traditional SAE and DBN. However, the average classification accuracies can achieve 74%~99% for fine-grained classification based on CNN combined with SVM, and the classification accuracies are better than the accuracies of 66%~98% in the existing literature. Compared with traditional CNN for 100 epochs, the training time of CNN combined with SVM can reduce about 90% at the same network structure.For the product image classification based on CNN combined with SVM, this thesis further studies the effect of classifying performance with 3 pooling methods(average pooling,max pooling, stochastic pooling). The results of experiment show that the classified performance of average pooling is best and much more stable with the CNN structure of 6layers or less, and max pooling also can achieve the same classification result as average pooling when the CNN structure is appropriate,what’s more,the classified performance of stochastic pooling is worst and has a certain degree of instability. For the CNN structure of 6more layers, max pooling is much more suitable, and the more the layers of CNN are, the better classification result of max pooling will be, but there has been a modest decrease in classification performance about average pooling, in addition, the classification results of stochastic pooling have a certain degree of randomness, but the law of experiment is close to max pooling. For the experiment with self-built product images of fine-grained classification and public images, the results show that the above law applies to both the fine-grained classification of product images and the classification of public images.