To recognize handwritten tax invoice text and manage tax invoices, a handwritten tax invoice recognition method based on improved differentiable binary network and convolutional recurrent neural network (CRNN) is proposed. This method first achieves bill text detection and localization through a differentiable binary network that integrates feature pyramid enhancement modules (FPEM). An improved CRNN is used to achieve bill text content recognition. In terms of text detection, the proposed module has an average detection precision of 89.1%, a recall rate of 86.4%, and an average frame rate of 63.5 frames per second. Compared with the Text Detection model based on Segformer and Enhanced Feature Pyramid and the Improved Fourier Contour Embedding Network, the precision has increased by 4.9% and 3.5%, respectively, and the frame rate has increased by 20.5 and 11.7 frames per second. In terms of text recognition, the improved CRNN has an average recognition precision of 95.6%, a recall rate of 94.2%, an average frame rate of 56.5 frames per second, and an F1-Score of 0.94. Compared with the Dual-Stream Network Fusing Spatial and Frequency Domain Feature and the Multi-modal Network based on Visual Attention and Semantic Perception, the precision has been improved by 4.9% and 3.0%, respectively, and the F1-Score has been improved by 0.07 and 0.04. This method performs excellently in processing complex handwritten tax receipts, with outstanding recognition precision and robustness, significantly reducing tax workload, and improving tax declaration and management efficiency.
Wei et al. (Tue,) studied this question.