Scene text segmentation based on redundant representation of character candidates

Abstract

Text segmentation is important step in extraction of textual information from natural scene images. This paper proposes novel method for generation of character candidate regions based on redundant representation of subpaths in extremal regions (ER) tree. These subpaths are constructed using area variation and pruned using their length: each sufficiently long subpath is character candidate which is represented by subset of regions contained in the subpath. Mean SVM probability score of regions in subset is used to filter out non character components. Proposed approach for character candidates generation is followed by character grouping and restoration steps. Experimental results obtained on the ICDAR 2013 dataset shows that the proposed text segmentation method obtains second highest precision and competitive recall rate.