Handwritten Character Recognition Based on Improved Convolutional Neural Network

Abstract

Because of the characteristics of high redundancy, high parallelism and nonlinearity in the handwritten character recognition model, the convolutional neural networks (CNNs) are becoming the first choice to solve these complex problems. The complexity, the types of characters, the character similarity of the handwritten character dataset, and the choice of optimizers all have a great impact on the network model, resulting in low accuracy, high loss, and other problems. In view of the existence of these problems, an improved LeNet-5 model is proposed. Through increasing its convolutional layers and fully connected layers, higher quality features can be extracted. Secondly, a more complex dataset called EMNIST is selected and many experiments are carried out. After many experiments, the Adam optimization algorithm is finally chosen to optimize the network model. Then, for processing character similarity problems on the pre-processed EMNIST dataset, the dataset is divided into different parts and to be processed. A better-divided result is selected after the comparative experiments. Finally, the high accuracy recognition of handwritten characters is achieved. The experimental results show that the recognition accuracy of the handwritten characters reached at 88% in the test set, and the loss is low.