Optimal Parameter Selection Technique for a Neural Network Based Local Thresholding Method

Abstract Thresholding of a given image into binary image is a necessary step for most image analysis and recognition techniques. In document recognition application, success of OCR mostly depends on the quality of the thresholded image. Non-uniform illumination, low contrast and complex background make it challenging in this application. In this paper, selection of optimal parameters for Neural Network (NN) based local thresholding approach for grey scale composite document image with non-uniform background is proposed. NN-based local image thresholding technique uses 8 statistical and textural image features to obtain a feature vector for each pixel from a window of size (2n + 1)x(2n + 1), where n ≥ 1. An exhaustive search was conducted on these features and found pixel value, mean and entropy are the optimal features at window size 3x3. To validate these 3 features some non-uniform watermarked document images with known binary document images called base documents are used. Characters were extracted from these watermarked documents using the proposed 3 features. The difference between the thresholded document and base document is the noise. A quantitative measure Peak-Signal-to-Noise ratio (PSNR) is used to measure the noise. In case of unknown base document characters were extracted through the proposed 3 features and used in a commercial OCR to obtain the character recognition rate. The average recognition rate 99.25% and PSNR shows that the proposed 3 features are the optimal compare to the NN-based thresholding technique with different parameters presented in the literature.