YAM: A Novel Binarization approach for Non-Uniformly Illuminated Document Images

Document image binarization is a crucial preporocessing step in Optical Character Recognition (OCR) systems, but its accuracy is often hindered by non-uniform illumination. This paper introduces YAM, a novel and computationally efficient binarization technique based on a hybrid thresholding strategy. YAM integrate both global image statistics and local neighbourhood analysis to robustly adapt to luminance variations. Experiments conducted on the WEZUT OCR dataset shows YAM demonstrated superior performance over traditional binarization methods, including Sauvola, Niblack, and Otsu, achieving an F-measure of text recognition of $97.54 \%$ and a word accuracy score of $97.74 \%$. YAM incorporates a box filter approach which makes it executes faster than recent binarization techniques like ZigZag. Both quantitative and qualitative results validate YAM offers a compelling balance between performance and efficiency as a practical and high-fidelity solution suitable for resource-constrained environments.