Stratification - based super - resolution algorithm for document processing

This paper presents a super-resolution algorithm specifically aimed at the document processing. There are some particular characteristics in document image, such as step edge, fixed pattern etc. And in the practical application scenarios, such as scanner optimization, electronic books and so on, all of them propose a real-time requirement to the algorithm. For these scenarios, the existing interpolation-based and learning-based super-resolution methods have their limitations. Based on the idea of stratification, this paper proposes the application of guided filter in document matting, and divides the image with this approach into three layers: foreground color, background color and text edge. Interpolated algorithm are used for foreground and background parts to maintain their smoothness and high fidelity of color. For the key structure information contained in the edge layer, the super-resolution reconstruction is performed using the training-based algorithm. The algorithm has the characteristics of color fidelity, edge preserving, clear structure, and efficient training, which can meet the application requirements of document processing well. This paper provides a detailed description of the algorithm, and at the same time provides sufficient comparison and quantification experiments to demonstrate the effectiveness of the algorithm.