Automatic classification and recognition of complex documents based on Faster RCNN

摘要

OCR(Optical Character Recognition) has been widely used in digital document processing, but the current OCR technology only works well in simple document processing. In contrast, complex documents contain a lot of non-text information (icons, forms, signatures, seals, noise, etc.). Most OCR systems often misinterpret these non-textual information as text when dealing with complex documents, resulting in the wrong target recognition. Automatic document segmentation and recognition technology not only improves the processing accuracy of OCR, but also enhances the processing of documents. Therefore, this project will study how to use Faster RCNN (fast regional convolution) technology to automatically classify and identify text, icon, table, noise and other objects in complex documents. Faster RCNN is one of the mainstream frameworks in the field of target detection. Although it has been applied to text area recognition, it has not been reported in the division and recognition of complex documents. This project specifically studies the complex document image preprocessing technology, RPN (Region Proposal Network) technology and target recognition technology involved in the Faster RCNN technology. It enables fast and accurate separation of seal area, text area and page number area from complex documents. It provides accurate target area for the next step of OCR recognition and document enhancement, and realizes accurate recognition and enhancement.