Text tracking and recognition from mobile device video

Abstract

Accuracy rates of OCR on the texts embedded in natural scene captured by a mobile device depends on quality of enhancement processing on image resolution degraded by poor illumination, motion blur, and distortion on the text containing surface due to the camera‘s angle of view. In this paper, we present various methods to improve text localization for mobile device captured texts in the natural scene. Simulations show that, even in the calibrated document acquisition environment, OCR accuracy on the surface distorted by perspective projection is very low. Surface re-projection is a standard preprocessing method for improvement of recognition. In this paper, we adopt the pinhole camera projection model to estimate the parameters for the perspective mapping. For the construction of patches of rectification mapping around the distorted text surfaces, we apply quad-tree construction by utilizing geometric structure of text paragraph, i.e., alignment of text baselines, to estimate the parameters of perspective map. © Research India Publications.