Text localization in natural scene images with stroke width histogram and superpixel

Abstract

A novel stroke-based method to localize text in unconstrained natural scene images is proposed. Firstly, in order to improve the edge detection in tough situations where the texts are partially occluded or noisy, we use stroke width histogram as guidance to generate a series of superpixels. Secondly, we present a novel way of using distance transform and sobel operator to extract character skeleton and then use the skeleton to improve stroke-width accuracy. Our method was evaluated on two standard datasets: ICDAR 2005 and ICDAR 2011, and the experimental results show that it achieves state-of-the-art performance.