Abstract
In this paper, a novel binarization technique is introduced for natural scene text, which can be applied after the text location step in order to improve OCR recognition. At the first step, an “optimum” conversion from color image to grayscale image is performed by minimizing L1 - Norm distance between original color image and reconstructed image on corresponding optimum projection vector. Based on it, at the second step, an approach is developed to classify scene text into two categories: “simple” and “complex”, for the purpose of optimizing the processing speed and preserving performance. At the last step, binarization is performed with different methods for “simple” and “complex” scene text. Results on word images from the challenging ICDAR 2003 dataset show that our scheme can gain higher performance compared with state-of-the-art methods in OCR accuracy.