Abstract

This paper presents a novel scene text detection algo-rithm, Canny Text Detector, which takes advantage of thesimilarity between image edge and text for effective text lo-calization with improved recall rate. As closely related edge pixels construct the structural information of an object,we observe that cohesive characters compose a meaning-ful word/sentence sharing similar properties such as spa-tial location, size, color, and stroke width regardless of lan-guage. However, prevalent scene text detection approacheshave not fully utilized such similarity, but mostly rely on thecharacters classified with high confidence, leading to low recall rate. By exploiting the similarity, our approach can quickly and robustly localize a variety of texts. Inspired by the original Canny edge detector, our algorithm makes useof double threshold and hysteresis tracking to detect textsof low confidence. Experimental results on public datasets demonstrate that our algorithm outperforms the state-ofthe-art scene text detection methods in terms of detection rate.