Text detection in scene images based on feature detection and tensor voting

Abstract

In this paper, a novel approach based on feature points to localize text in natural scene images is proposed. The key idea of this approach is: feature point detection technique is used to extract the corner points of edges in connected components, to select candidate regions. Then, the candidate regions are verified by tensor voting, to extract perceptual structures from noisy data. Finally, non-text regions are filtered out by using regions area, orientation, and aspect ratio. The experiment results demostrate the performance of our approach.