FEASIBILITY STUDY OF CROSSWISE REGION MERGING FOR SCENE TEXT LOCALIZATION WITH TWO-CLASS DETECTOR

摘要

Scene text recognition is a challenging research topic in the field of image recognition. To solve this problem, in addition to derived schemes from traditional OCR techniques for scanned document, a novel approach based on generic object detection scheme whose detection accuracy has been improved drastically in recent years is becoming popular. This paper proposes a novel text localization scheme aiming to improve both recognition accuracy and processing speed of scene text recognition. The proposed scheme adopts two-class detector for initial character detection and region merging for text localization to execute proper preprocessing before character recognition by multi-class classifier. Experimental results using ICDAR dataset have shown that the proposed scheme can reduce search area to 36.4% while false negative rate is 10.8%. By the results, it is shown that the proposed scheme is feasible for scene text recognition.