Global Binarization of Document Images Using a Neural Network

Abstract

In degraded scanned documents, where considerable background noise or variation in contrast and illumination exists, pixels may not be easily classified as foreground or background pixels. Thus, the need to perform document binarization in order to enhance the document image by separating foregrounds (text) from backgrounds. A new approach that combines a global thresholding method and a supervised neural network classifier is proposed to enhance scanned documents and to separate foreground and background layers. Thresholding is first applied using mass-difference thresholding to obtain various local optimum threshold values in an image. The neural network is then trained using these values at its input and a single global optimum threshold value for the entire image at its output. Compared with other methods, experimental results show that this combined approach is computationally cost effective and is capable of enhancing degraded documents with superior foreground and background separation results.