Parallel nonparametric binarization for degraded document images

Adaptive binarization has been widely used in binarizing degraded document images. Most of the adaptive methods, however, face two challenging problems: expensive computation and sensitivity to introduced parameters. To solve the two challenges, we propose a novel parallel nonparametric method consisting of three steps: (i) achieving a number of binary images using Sauvola׳s method with different parameters; (ii) recognizing each pixel of these binary images using linear SVMs, and (iii) reconstructing a binary image on the basis of the recognized binary images. Our method therefore is a new concept to binarize an image. Instead of computing appropriate thresholding values, we generate a new binary image in term of numerous recognized binary images. The prerequisite of this idea is big enough data generated, and the modern CUDA-enabled GPUs provide the powerful computation capacity. In our work, we develop a CUDA well-suited parallel algorithm of Sauvola׳s method and implement it on Kepler GPUs with CUDA 5.0. Overall, our proposed method is highly parallelized as well and easily implemented on distributed systems if higher performance required. Experimental results on four public challenging datasets have shown that our proposed method outperforms the state-of-the-art methods.