The Enhancement of Character and Non-Character images using Extensive segmentation Techniques

Abstract

The research work presents detection of different types of text in scene images based on extensive segmentation (proposed method) to generate character candidate regions. We usually consider many connected regions as candidates, which aim to capture character regions as many as possible. Key feature of exhaustive segmentation technique, which exactly segments character candidate region in scene images from non-character candidate region also. First, we detect candidate text regions using Maximally Stable External Region (MSER method) where the scene image has been converted to gray image and to find the text region. Second, the geometric properties of text on the image are used to filter out non-text regions using simple thresholds. Third, we remove the rest Non-Text region based on Stroke Width Variation (SWV). Finally, we merge the entire text region from the detection of Text and thus recognise the Detected Text in the scene image. We use public dataset, namely, the Street View Text dataset and some other language (Tamil, Hindi and Chinese) images to evaluate the performance of our(extensive segmentation) method. The experimental results are shown that our(extensive segmentation) method achieves excellent improvement in the detection of text though the images being blurred, low-resolution and small in size from the existing method(Yuenwang et al.). We also achieve considerable rate of recall with the executed images.