Research and realization of commodity image retrieval system based on deep learning

This paper proposed a commodity image retrieval system based on CNN and ListNet sort learning method. CNN contained two convolutional layers, two pooling layers and two innerproduct layers. ReLu function was used as the activation function after the convolutional layer, achieving the sparsity and preventing the disappearance of the gradient. The pooling layer used stochastic pooling method and improved the generalization ability of the model. In addition, softmax regression was used for classification. Innerproduct layer adopted dropconnect method, which is more powerful than the generalization of dropout, and it can effectively prevent the occurrence of the overfitting. What’s more, the feature extraction of the network was optimized by stochastic gradient descent (SGD) algorithm. And we combined the learn to rank algorithm of the text retrieval domain. We used ListNet algorithm to combine a variety of feature vectors, solving the problem of the image retrieval.