Handwritten digit recognition based on improved convolution neural network

Abstract

An improved convolution neural network model is proposed, which has higher recognition rate for handwritten digits. Based on the AlexNet network model, the algorithm improves the feature extraction ability of the model by introducing residual module to modify the third and fourth volume layers in the model. The batch normalization (BN) method is used to prevent over-fitting after each convolution. In order to reduce the amount of computation, a full connection layer is reduced. The algorithm has a good effect on handwritten digit recognition by training and testing on MNIST dataset. Compared with AlexNet network model, the improved model has higher detection accuracy.