Natural scene text localization and detection using MSER and its variants: a comprehensive survey

Text localization and detection within natural scene images have generated significant interest among researchers due to their inherent complexity and various real-life applications. In the last few decades, various methodologies have been developed for localization and detection of wild scene text regions. Among them, Maximally Stable Extremal Regions (MSER) based techniques have achieved remarkable success in a significant variety of text localization tasks over the last decade. MSER is a well-known blob detection method, which has been applied with some modifications in many scene text-related researches. In this paper, we have reviewed and evaluated the concept of MSER methods which are combined with traditional machine learning-based methods using hand-crafted features or deep learning-based methods using automatic feature learning for scene text localization. Different MSER methods, such as standard MSER, MSER with stroke width transform, eMSER, enhanced MSER, multi-level MSER, MSER with CNN features, component splitting with MSER tree, MSER with CNN and CRF, CE-MSER have been described in this study. Finally, we have compared and evaluated the performances of those different types of MSER methods on five publicly available standard scene text datasets, like ICDAR 2003, ICDAR 2013, ICDAR 2015, KAIST, and SVT and provided the insights of appropriate selection of MSER method along with its pros and cons.