Parameter Tuning in MSER for Text Localization in Multi-lingual Camera-Captured Scene Text Images

Scene text detection and localization in camera-captured images has always posed a great challenge to the researchers due to its high complexity in understanding the texture and homogeneity of scene text images. The solution to this problem paves way for simplified text extraction and processing, thereby realizing wide range of applications. A very popular method, namely, maximally stable extremal region or MSER detection is used for localizing text since the text regions are considered to be more stable than other regions in an image. However, it involves manual selection of several parameters, limiting its usage in many practical applications. In this work, the relations among parameters, like image dimension, text size, and region area, are analyzed by experimenting on an in-house multi-lingual dataset having 300 images comprising English, Bangla, and Hindi texts and validating the outcome against images of standard datasets, like SVT and MSRA-TD500, which yields encouraging results.