Fractals based multi-oriented text detection system for recognition in mobile video images

Text detection in mobile video is challenging due to poor quality, complex background, arbitrary orientation and text movement. In this work, we introduce fractals for text detection in video captured by mobile cameras. We first use fractal properties such as self-similarity in a novel way in the gradient domain for enhancing low resolution mobile video. We then propose to use k-means clustering for separating text components from non-text ones. To make the method font size independent, fractal expansion is further explored in the wavelet domain in a pyramid structure for text components in text cluster to identify text candidates. Next, potential text candidates are obtained by studying the optical flow property of text candidates. Direction guided boundary growing is finally proposed to extract multi-oriented texts. The method is tested on different datasets, which include low resolution video captured by mobile, benchmark ICDAR 2013 video, YouTube Video Text (YVT) data, ICDAR 2013, Microsoft, and MSRA arbitrary orientation natural scene datasets, to evaluate the performance of the proposed method in terms of recall, precision, F-measure and misdetection rate. To show the effectiveness of the proposed method, the results are compared with the state of the art methods.