Based on the Inception and the ResNet Module Improving VGG16 to Classify Commodity Images

Abstract

With the advent of the new retail era, intelligent identification of goods in the shelf image has become an important technology for managing unmanned supermarkets. In recent years, classification methods based on convolutional neural networks have been widely used in image classification. In this paper, a deep neural network based on VGG16-IR is designed to improve the classification accuracy of low-resolution commodity images. First, Inception and residual ideas are integrated, and four modules are designed, namely Inception ResNet module 1, Inception ResNet module 2, ResNet module 3, and ResNet module 4. These modules replace the convolution phase of the original VGG16, and VGG16-IR replaces the new network. Secondly, a small batch gradient descent algorithm with momentum is used. Finally, the activation function of ELU is used to avoid the phenomenon of partial neuron necrosis. The experiment is conducted in Microsoft PI100 product image training and testing. The experimental results show that the model's accuracy in the training set reaches 99.69%, and the accuracy of the test set reaches 90.63%. Compared with 89.10% of VGG16, the model in this paper improves 1.5% and increases the generalization ability to a certain extent.