Abstract
The evaluation of natural scene images for text localization is an appealing task to examine the image contents. In this paper, MSER-based candidate character regions are initially compared with the geometric features based effective text localization method based to find text regions. In addition, zone-based features of MSER-based extracted complementary candidate characters are computed for respective zones including the regional features. Bayesian logistic regression classifier is trained on features complementary candidate characters. The complementary candidate character regions with higher posterior probability are considered as candidate characters or letters corresponding to non-candidate characters or letters. Adjacent complementary candidate characters with higher posterior probabilities are grouped into words and sentences. Consequently, zone-based text localization, named as ZText, is evaluated on ICDAR 2015 Robust Reading Competition benchmark dataset. The results of experiments have established amazing competitive performance with the recently published text localization algorithms.