Local thresholding of composite documents using multi-layer perceptron neural network

Bi-level thresholding of document images with poor contrast, non-uniform illumination, complex background patterns and non-uniformly distributed backgrounds is a challenging problem that researchers have been trying to solve. The problem is that different algorithms tend to yield different results based on the assumptions made to the images content. A new binarization algorithm is proposed to deal with such images. The algorithm proposed uses statistical and texture feature measures to obtain a feature vector from a pixel window of size (2n+1)/spl times/(2n+1), it then uses a multi-layer perceptron neural network (MLP NN) to classify each pixel value in the image. The proposed method performed better than existing global and local thresholding techniques and works on different variety of images. The algorithm provides a local understanding of pixels from its neighborhood. This new method that uses NN and works on scanned documents with non-uniform backgrounds.