A PDE system based on background estimation for document binarization with non-uniform background

Abstract

Document binarization, despite extensive research, still remains a challenging task for images with non-uniform background (NUB). Background estimation (BE) is an effective preprocessing technique to cope with this challenge. Previous BE-based binarization estimates a background component from the input and uses that estimate to preprocess the input to yield the compensated image, followed by a commonly sophisticated processing step for binarization. In this paper, inspired by BE, we introduce a binarization model in the framework of partial differential equation (PDE) for NUB document images. Specifically, we first model the NUB degradation as an uneven scaling map that adjusts the contrast of a normal document to produce the NUB document image. In other words, a NUB document image can be represented as the pixel-wise product of the uneven scaling map and the normal image (i.e., compensated image). The problem of compensating for the background is therefore rendered into the restoration problem of recovering the normal document image from the input NUB document image. Based on the NUB model, we propose a PDE system for binarization of document images with NUB degradation, which consists of an evolution PDE (w.r.t the scaling component) with the input NUB image as the initial condition, and a nonlinear diffusion equation (w.r.t the compensated component) with a pure white image as the initial condition. This PDE system estimates the scaling component and the compensated image dynamically and alternately during evolution. In the final compensated image, the text and background pixels are divided into two primary modes with a separation of 0.5, thereby enabling us to achieve binarization of the input NUB document image just by a binary projection separated by 0.5. Finally, the PDE system is evaluated on twenty-nine NUB document images from publicly available DIBCO datasets, in comparison to seven related PDE binarization models. Experimental results show that the proposed PDE system outperforms the seven models for comparison in terms of binarization of NUB document images.