Character Recognition in Images under High Noise Levels
Alexander Konanykhin, Tatyana Konanykhina, В. С. Панищев · 2023
The task of extracting information in the form of text from scanned or photographed objects using Optical Character Recognition (OCR) is an important and ubiquitous technology for digitizing and indexing physical documents. Existing technologies have shown to work correctly with near-ideal documents, but if the document is visually degraded or contains non-text elements, the quality of OCR can be greatly affected, especially due to false detections. In this article, we present an improved detection network with a masking system to improve the quality of document recognition. By filtering out non-text elements in an image, we can use document-level OCR to include contextual information to improve OCR results. We conduct a single evaluation of a publicly available data set, demonstrating the usefulness and broad applicability of our method. In addition, we present data and calculation results of a categorical cross entropy function specially tuned to improve data discovery results.