MSdB-NMF: MultiSpectral Document Image Binarization Framework via Non-negative Matrix Factorization Approach.

Area of Science:

• Computer Vision and Image Processing
• Machine Learning
• Digital Image Analysis

Background:

• Multispectral document image binarization (MSdB) is crucial for document analysis but faces challenges due to spectral variations and noise.
• Existing methods often require prior knowledge of spectral bands or lack robustness against various image degradations.

Purpose of the Study:

• To propose a novel, robust, and adaptable framework for Multispectral document image binarization (MSdB) using Non-negative Matrix Factorization (NMF).
• To develop an NMF-based feature extraction method that enhances robustness and sparseness for improved binarization performance.

Main Methods:

• A three-step MSdB-NMF framework involving NMF-based feature extraction with a new optimization problem, a post-processing selection method, and application of existing binarization schemes.
• Extraction of N features from B spectral bands (N < B) using multiplicative updating rules and convergence proof for the feature extraction algorithm.
• Feature vector selection via visual inspection or an automated post-processing method using benchmark binarization results.

Main Results:

• The proposed MSdB-NMF framework demonstrated superior performance compared to several state-of-the-art binarization methods on two MS document image datasets.
• The method achieved competitive results, outperforming the winner of the MS-TEx-2015 contest, highlighting its effectiveness and robustness.
• The framework is applicable to any MS or hyperspectral (HS) document image without requiring prior spectral band information.

Conclusions:

• The novel MSdB-NMF framework offers an effective and versatile solution for multispectral and hyperspectral document image binarization.
• The proposed NMF-based feature extraction and selection strategy significantly enhances binarization accuracy and robustness against image degradations.