Scene Text Extraction with Edge Constraint and Text Collinearity

Abstract

In this paper, we propose a framework for isolatingtext regions from natural scene images. Themain algorithm has two functions: it generates textregion candidates, and it verifies of the label ofthe candidates (text or non-text). The text regioncandidates are generated through a modified Kmeansclustering algorithm, which references texturefeatures, edge information and color information.The candidate labels are then verified in aglobal sense by the Markov Random Field modelwhere collinearity weight is added as long as mosttexts are aligned. The proposed method achievesreasonable accuracy for text extraction from moderatelydicult examples from the ICDAR 2003database.