A Comparative Study for Unconstrained Word Detection Algorithms in Natural Images
Abstract
Detecting text in unstructured scenes is a challenging task because of multi-orientation, perspective distortion, and variation of text size, color and scale. The problem becomes more difficult when the multilingual text in different orientations is encountered. In this paper, we describe the detection of text from natural images which includes images with text having different languages, different orientation (vertical, horizontal, etc.), different styles (stray images, text shape etc. Following paper analyses different existing techniques to detect text from different images including different orientations and styles. We have tested these techniques on standard datasets that are as follows-ICDAR 2017, MSRA-TD 500, and Personalised Dataset. We apply various detection algorithms, a combination of exhaustive search and segmentation. Five different existing algorithms (MSER, Otsu, Edge Based Text Region Extraction, Connected Component Based Text Region Extraction, and Saliency) are analysed and measured over standard performance parameters.
Keywords
MSER, OTSU, Text Detection, Segmentation