In order to improve the accuracy and efficiency of paper financial reimbursement vouchers, a deep learning based automatic recognition strategy for financial accounting vouchers is proposed. This article designs the basic process of deep learning for image classification and text recognition, and establishes invoice image datasets for training CNN networks. Next, parameter design is performed on the AlexNet network structure, and the trained network model is used to improve the digital recognition method. The experimental part of the text annotation system collected multiple invoice text recognition datasets, and the optimized model is subjected to network training and testing experiments on the dataset. The test results show that the detection accuracy of specific fields in the invoice image reaches over 95%, and the single item recognition rate of each field reaches about 95%. The model accuracy and inference speed can meet the performance requirements.