Abstract
Handwritten digit recognition has been widely researched. Numerous classification methods have been proposed. This article improves the recognition of handwritten numbers method based on CNN neural network. Recognition of handwritten digits has made great progress, but due to insufficient recognition accuracy, it will have a certain impact on work efficiency. In this paper, a two-layer CNN network with two fully connected layers can achieve a good prediction performance. In our framework, the activation function used is the ReLU function, to some extent, this avoids the difficulties of gradient disappearance and gradient saturation. At the same time, dropout is added to the full connection layer to discard the neurons in the network with a certain probability, which can avoid certain accidents, and the final trained network will be more robust and the generalization performance will be stronger. This method effectively improves the accuracy of recognition in the final experiment, which achieves 98.75% accuracy on theMNIST dataset and 99.25% on the CIFAR-100 dataset.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call