Scene text detection via extremal region based double threshold convolutional network classification.

Abstract

In this paper, we present a robust text detection approach in natural images which is based on region proposal mechanism. A powerful low-level detector named saliency enhanced-MSER extended from the widely-used MSER is proposed by incorporating saliency detection methods, which ensures a high recall rate. Given a natural image, character candidates are extracted from three channels in a perception-based illumination invariant color space by saliency-enhanced MSER algorithm. A discriminative convolutional neural network (CNN) is jointly trained with multi-level information including pixel-level and character-level information as character candidate classifier. Each image patch is classified as strong text, weak text and non-text by double threshold filtering instead of conventional one-step classification, leveraging confident scores obtained via CNN. To further prune non-text regions, we develop a recursive neighborhood search algorithm to track credible texts from weak text set. Finally, characters are grouped into text lines using heuristic features such as spatial location, size, color, and stroke width. We compare our approach with several state-of-the-art methods, and experiments show that our method achieves competitive performance on public datasets ICDAR 2011 and ICDAR 2013.