基于TrOCR的中文手写文本识别研究OA

Chinese handwritten text recognition has important application value in document digitization.TrOCR,as an end-to-end recognition model based on Transformer,performs well in handwritten text recognition tasks,but it cannot directly process Chinese characters using an English vocabulary.The article focuses on the Chinese handwriting recognition task and reconstructs the decoder Chinese vocabulary on the CASIA-HWDB2 dataset to adapt TrOCR to the Chinese recognition task while retaining the feature extraction ability of the pre-trained visual encoder.On this basis,carry out comparative experiments to systematically evaluate the impact of data augmentation and Dropout regularization on performance.The experimental results show that based on the baseline of pre-trained model character accuracy of 98.11%and sequence accuracy of 81.20%,after introducing data augmentation and Dropout regularization,the final model character accuracy reaches 98.89%and sequence accuracy reaches 84.80%.The research results provide a basis for TrOCR adaptation and training strategy optimization in Chinese handwritten text recognition.