Offline Handwritten Chinese Recognition Based on Multi-Branch Convolution and Lightweight ECA
Ling Sun, Chengyu Wen, Gong Chen, Binlin Zhao · 2024
Chinese handwriting recognition technology, as a key link in converting Chinese handwritten text into electronic format, has become increasingly valuable in research and application. Due to the variety of Chinese language, complex structure, and shape similar characters, and different writers writing arbitrariness, offline handwritten Chinese characters lack of stroke order support and other issues, resulting in the lack of expressive power of feature extraction has been a major obstacle to the improvement of the recognition accuracy. In order to solve this problem, this paper proposes an offline handwritten Chinese recognition method based on multi-scale branching convolution and lightweight attention mechanism, which is added to the current mainstream CRNN+CTC text recognition framework. First, the basic features of text images are extracted by a convolutional neural network constructed by Inception module and lightweight attention mechanism ECA; then, the extracted features are predicted using recurrent neural network and a probability distribution about the character set of Chinese characters is outputted; finally, the recognition results are computed by using a connectionist sequence classification algorithm and a loss function is constructed. Experiments are conducted on the CASIA-HWDB2.0-2.2 dataset of handwritten Chinese character texts using the proposed method, and the results show that the method can obtain 96.720/0 accuracy, which proves that the method is feasible.