Accounting-Oriented Research on Note Recognition Model based on Information Extraction Algorithm

Enterprise accountants deal with bill reimbursement mostly relying on the traditional manual way to carry out, and the current bill recognition technology makes it difficult to meet the recognition needs of Chinese bills. And there is a lack of open-source Chinese bill recognition models in the training and validation process of the billing model. Aiming at the above challenges, the study proposes an information extraction algorithm based on the optical character recognition technique of deep learning, and the bill recognition model construction is carried out on this basis. Image detection is performed by utilizing detection and recognition neural networks, and image feature extraction is performed by combining convolutional recurrent neural networks with connectionist temporal classification. The validation shows that the accuracy of the research-proposed information extraction algorithm increases by an average of 9.86% compared with other algorithms in the self-constructed cab invoice dataset, and the F1 value in the International Conference on Integration and Innovation of Digital Archival Resources Toward the Enhancement of Public Service Capability 2015 dataset increases by 5.82% and 0.92% compared with other algorithms, respectively. Compared to other models, the study’s proposed model increases the average number of frames per second by 34.47% and the average class-wide accuracy by 10.72% in the cab invoice dataset. The bill recognition model based on the information extraction algorithm proposed in the study can meet the bill recognition requirements, has superior recognition accuracy and efficiency, and has application value in enterprise bill recognition.