Unsupervised document image binarization using a system of nonlinear partial differential equations

Abstract

Partial differential equations have recently been established as a viable framework for image processing, particularly for image binarization. One drawback of this framework is the requirement for manual parameter tuning. In this work we propose a novel development wherein the spatio-temporal dynamics of the thresholding parameter are governed by an additional partial differential equation which is engineered to exhibit desirable traits. While the model can still be tuned manually to achieve optimal results, we show experimentally that the present framework is near optimal for the default choice of parameter, τ. This novel system enforces a smooth evolution of the threshold map while still offering locally adaptive thresholding properties, a requirement for non-uniformly illuminated images. The proposed model is applied to images through a rudimentary finite difference based numerical method due to the parallelizability and provable stability of the method. The proposed work offers an unsupervised binarization scheme and is benchmarked against state-of-the-art methods in the field.