2021/05/10 by Shachar Klaiman, Klaiman, Shachar, Marius Lehne +1
Computer Science · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Handwritten Text Recognition Techniques #Image Processing and 3D Reconstruction #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2105.04313
openalex publication_date 2021/05/10 · openalex created_date 2022/09/16 · openalex updated_date 2026/07/28
Information extraction from documents is a ubiquitous first step in many\nbusiness applications. During this step, the entries of various fields must\nfirst be read from the images of scanned documents before being further\nprocessed and inserted into the corresponding databases. While many different\nmethods have been developed over the past years in order to automate the above\nextraction step, they all share the requirement of bounding-box or text segment\nannotations of their training documents. In this work we present DocReader, an\nend-to-end neural-network-based information extraction solution which can be\ntrained using solely the images and the target values that need to be read. The\nDocReader can thus leverage existing historical extraction data, completely\neliminating the need for any additional annotations beyond what is naturally\navailable in existing human-operated service centres. We demonstrate that the\nDocReader can reach and surpass other methods which require bounding-boxes for\ntraining, as well as provide a clear path for continual learning during its\ndeployment in production.\n