부산대학교 도서관

In this paper, we propose a deep learning base batch invoice recognition system, obtaining highly accurate recognition for value-added tax (VAT) invoice. Invoices are scanned in batch and imported into computer to derive raw images, with which our work proceeds in the following steps: 1) raw images are normalized; 2) we improve the character segmentation by the projection algorithm according to location information in template matching algorithm; 3) convolutional neural network (CNN) was further adopted for character recognition; 4) we introduce a module for judgment and manual modification, which estimates recognition considering the association of invoice contents, and then the wrong recognized invoice is modified to improve the system's fault tolerance; 5) the correct information is recorded. Empirical results show that the location algorithm based on template matching is most suitable for the invoice system among several tested algorithms. The application of deep learning technology to character recognition ensures the recognition accuracy. The performance of the entire system was also evaluated with random invoices. The proposed system, with the aid of the manual modification, can reach 100% recognition rate.