Text Localization in Natural Images Using Multiple Feature Fusion

Abstract

Text localization in natural scene images is an important first step to analyze content-based images. In this thesis, we propose an accurate method for detecting texts in natural scene images by adapting multiple feature combinations. Firstly, various color spaces are used to extract Maximally Stable Extremal Regions (MSERs) as character candidates. K-means clustering, image gradients, and Lucy and Richardson de-convolution (LRD) method are used to emphasize significant feature points and to compensate noisy images for character candidates. Secondly, important character candidates will be accentuated by applying enhanced canny edge detector to grayscale image gradients and our cumulative MSERs image. Thirdly, character candidates will be merged into text candidates by using clustering algorithm and geometric information such that texts usually appear in a linear form. Finally, inconsequential text candidates will be eliminated by using stroke width, text region height, width, and parallel edge features. The method was evaluated on two benchmark datasets: International Conference on Document Analysis and Recognition (ICDAR) 2013 and Street View Text (SVT) from Google Maps. Experimental results on respective datasets show that the proposed algorithm works successfully with not only the normal text images, but also highlighted, transparent, small, and blurred texts.