关键词: 手写汉字识别, 卷积神经网络, 特征分组, 信息精炼和浓缩

Abstract:

For traditional offline handwritten Chinese character recognition, the process is complex and the accuracy is low. The feature information extraction of common convolutional neural networks is insufficient, and there are overlap and redundancy problems of the same feature information. In this paper, a deep convolutional neural network model for feature group extraction and fusion is designed. Through the multi-level stacked feature group extraction module, the deep abstract feature information of the image is extracted, and the communication and fusion between the feature information is performed. Using the designed downsampling and channel amplification module to preserve the important information of the image while reducing the feature dimension. The feature information is refined and condensed to solve the problem of overlapping and redundancy. In the end, the trained neural network reaches the current advanced accuracy of top1 （97.16%） and top5（99.36%）, with good generalization ability.

Key words: handwritten Chinese character recognition, Convolutional Neural Network（CNN）, feature grouping, information refining and concentration