Image Classification Based On improved VGG Network

Image classification is a fundamental problem in the field of computer vision. In this paper, we propose an improved VGG-16 model suitable for small data sets such as cifar10. The model reduces the fully connected layer, adds batch normalization (BN) and a stronger regularizer (Dropout) as well as improving the loss function. For the cifar10 data set, this paper uses four data augmentation methods: random flip, random translation, random erasure, and mix up to improve the generalization ability of the model. Experimental results show that the training parameters of the improved VGG model are reduced to 10.65% of the original model, and the average classification accuracy rate is increased to 94.43%.