Character Recognition Based On Deep Learning

Deep learning is a special algorithm of machine learning. Deep learning has made a series of breakthrough progresses in the field of image classification, object detection, image segmentation, speech recognition, etc. Covolutional Neural Network(CNN) is a significant network structure of deep learning. CNN is a kind of bionic network, namely it imitates biological neural networks, and it has some excellent properties that other detection or classification algorithms can't achieve. Character recognition also has wide applications in recent years. Chinese character recognition is a kind of tough and complicated issue in the field of character recognition.This article adopts deep convolutional neural networks and its variants for Chinese handwritten character recognition, details are as follows:1. Design deep covolutional neural networks, combined with the features of Chinese handwritten characters. The input of the network is Chinese handwritten character images, the output is classification result. The network parameters are updated through forward propagation training, back propagation training and gradient descent algorithm. In the comparative experiments, train and analyze different covolutional neural network structures so that we can study the influence of them on Chinese character recognition.2. Improve the structure of traditional deep convolutional neural networks by fusioning different deep convolutional neural network models, the improved network is called Multi-column deep neural network. This article proposed a probability average fusion algorithm of the output of Softmax regression model for Chinese handwritten character recognition, then the algorithm is applied to Multi-column deep neural network. The average fusion algorithm reduces the error rate of each covolutinal neural network of the Multi-column deep neural network, and then improved the accuracy of the whole network.3. Improve the training algorithm of traditional deep neural networks by decomposing the output features of covolutional layers into batches. Take batch normalization for each batch and then input them to next layer. This article proposes a network structure of GoogLeNet deep neural network combined with batch normalization algorithm for Chinese character recognition issue. The accuracy of traditional neural network was improved both by network structure and training algorithm.