Text detection in born-digital images by mass estimation

There is a need for effective web-document understanding due to the explosive progress of internet and network technologies. In this paper, we propose a new method for text detection in born-digital images by introducing a mass estimation concept. We propose to explore super-pixel information of different color channels to identify text atoms in images. The proposed method uses similarity graphs and spectral clustering to identify candidate text regions. We propose a new idea of mapping Gabor responses of a candidate text region to a spatial circle to study the spatial coherency of pixels. We introduce a mass estimation concept to identify text candidates from the pixel distribution in a spatial circle. The linear linkage graphs help in grouping text candidates to obtain full text lines. The same Gabor responses are used as features to eliminate false positives with an SVM classifier. We evaluate the proposed method for the testing on standard datasets, such as ICDAR 2013 (challenge-1) and the Situ et al. dataset. Experimental results on both the datasets show that the proposed method outperforms the existing methods.