Variation of Stability Factor of MSERs for Text Detection and Localization in Natural Scene Image Using Naive Bayes Classifier

The process of extracting textual regions from the scene images is a significant matter in the field of image processing & computer vision. It is very challenging due to different fonts, variable font size, illumination conditions and complex background etc. In last decade, image segmentation using Maximal Stable Extremal Regions (MSERs) played an important role in this area due to its various advantages. The generation of MSERs is controlled by variation of stability factor delta in deciding the promising stable areas. The aim of this paper is to study the effect of parameter delta and calculate the optimal delta on the different versions of MSER for detection and localization of text in scene images. Four different features Stroke Width Heterogeneity, Perpetual Color Contrast, Histogram of Oriented Gradients at Edges, Occupy Rate are used to evaluate the probability of text using naive Bayes Model for each version of MSERs. The Training is accomplished on the ICDAR 2013 training dataset and experiments for testing our method are carried out on ICDAR datasets to show the importance of delta (optimal value) parameter of MSER in providing the optimum results expressed as f-measure, recall and precision.