Pre-Processing Images of Public Signage for OCR Conversion

摘要

In this paper, we propose a novel method to enhance the OCR (Optical Character Recognition) readability of public signboards captured by smart-phone cameras—both outdoors and indoors, and subject to various lighting conditions. A distinct feature of our technique is the detection of these signs in the HSV (Hue, Saturation and Value) color space, done in order to filter out the signboard from the background, and correctly interpret the textual details of each signboard. This is then binarized using a thresholding technique that is optimized for text printed on contrasting backgrounds, and passed through the Tesseract engine to detect individual characters. We test out our technique on a dataset of over 200 images taken in and around the campus of our college, and are successful in attaining better OCR results in comparison to traditional methods. Further, we suggest the utilization of a method to automatically assign ROIs (Regions Of Interest) to detected signboards, for better recognition of textual information.