Handwritten Recognition Based on Convolutional Neural Networks

In recent years, with the rapid development of deep learning, convolutional neural networks(CNN) have been widely used in solving computer vision tasks. Handwriting recognition, as a subfield of pattern recognition, aims to recognize the content of an image and output it as text. Although CNN has excellent feature extraction ability, the letters that make up a word are related to each other, and CNN cannot use such spatial timing features. This paper uses the attention mechanism to increase CNN's ability to perceive local features and build the ACNN model. Using the frequency of each letter as prior knowledge, the FCNN model is established, and two branches are developed in the CNN model: evaluate after each epoch and use the improved model when the accuracy of FCNN handwriting recognition is higher than that of pure CNN model; When the accuracy of FCNN model is lower than that of CNN model, cnn model is used. The comparison experiment on the self-built French handwritten dataset ‘Yijuan’ shows that the recognition accuracy of the ACNN model is the highest, reaching 0.9889.