Improvement of mnist image recognition based on cnn

Abstract

At present, great progress has been made in the field of image recognition, especially in convolutional neural network. Lenet-5 convolutional neural network has been able to identify handwritten digit MNIST database with high precision. In this paper, experiments show that different activation functions, learning rates and the addition of the Dropout layer in front of the output layer will make the convergence speed different, weaken the influence of the initial parameters on the model, and improve the training accuracy. It is proved that the modified LeNet-5 model has a better improvement in handwritten digit recognition. This method is an efficient recognition method.