Scene text localization using extremal regions and Corner-HOG feature

Abstract

This paper presents a text detection method based on Extremal Regions (ERs) and Corner-HOG feature. Local Histogram of Oriented Gradient (HOG) extracted around corners (Corner-HOG) is used to effectively prune the non-text components in the component tree. Experimental results show that the Corner-HOG based pruning method can discard an average of 83.06% of all ERs in an image while preserving a recall of 90.51% of the text components. The remaining ERs are then grouped into text lines and candidate text lines are verified using black-white transition feature and the covariance descriptor of HOG. Experimental results on the 2011 Robust Reading Competition dataset show that the proposed text detection method provides promising performance.