[Paper Review] Degraded Historical Documents Images Binarization Using a Combination of Enhanced Techniques

Omar Boudraa, Walid Khaled Hidouci+1 more|arXiv (Cornell University)|Jan 27, 2019

Handwritten Text Recognition Techniques3 references4 citations

TL;DR

This paper proposes a three-stage hybrid binarization framework for degraded historical documents that combines CLAHE for contrast enhancement, a fusion of Otsu’s, multilevel Otsu’s, and Nick’s thresholding methods using local contrast as a decision criterion, and morphological post-processing for noise removal. The method achieves state-of-the-art performance on DIBCO and H-DIBCO benchmarks, with F-measure scores of 87.23% (DIBCO 2013), 92.40% (H-DIBCO 2014), and 85.08% (H-DIBCO 2016), outperforming Otsu, Sauvola, and Niblack methods.

ABSTRACT

Document image binarization is the initial step and a crucial in many document analysis and recognition scheme. In fact, it is still a relevant research subject and a fundamental challenge due to its importance and influence. This paper provides an original multi-phases system that hybridizes various efficient image thresholding methods in order to get the best binarization output. First, to improve contrast in particularly defective images, the application of CLAHE algorithm is suggested and justified. We then use a cooperative technique to segment image into two separated classes. At the end, a special transformation is applied for the purpose of removing scattered noise and of correcting characters forms. Experimentations demonstrate the precision and the robustness of our framework applied on historical degraded documents images within three benchmarks compared to other noted methods.

Motivation & Objective

• To address the persistent challenge of binarizing degraded historical document images with non-uniform illumination, noise, blur, and low contrast.
• To improve binarization accuracy by combining multiple thresholding techniques in a hybrid framework.
• To enhance image quality through preprocessing (CLAHE) and post-processing (connected components and morphological operations).
• To evaluate the method’s robustness across diverse degradation types using standardized benchmarks.
• To demonstrate superior performance compared to established global and adaptive thresholding methods.

Proposed method

• Preprocessing uses Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance contrast in low-contrast degraded document images.
• Hybrid binarization applies Otsu’s, multilevel Otsu’s, and Nick’s thresholding methods, with local average contrast used as a decision criterion to select the optimal threshold per region.
• The decision criterion selects the threshold from the three methods that maximizes inter-class variance while considering local contrast variations.
• Post-processing employs connected components analysis and morphological operations (e.g., opening and closing) to remove isolated noise and correct character shapes.
• The framework integrates three distinct thresholding algorithms in a cooperative manner, avoiding reliance on a single method’s limitations.
• Parameter tuning is performed empirically across eleven parameters, with future work suggesting automated tuning via fuzzy logic or deep learning.

Experimental results

Research questions

• RQ1 Can a hybrid combination of global and adaptive thresholding methods improve binarization accuracy on severely degraded historical documents?
• RQ2 Does CLAHE preprocessing significantly enhance the performance of subsequent thresholding techniques in low-contrast document images?
• RQ3 How does the proposed method compare to established methods like Otsu, Niblack, Sauvola, and Nick in terms of F-measure, DRD, and PSNR across diverse benchmarks?
• RQ4 To what extent does post-processing with morphological operations and connected components reduce noise and improve character integrity?
• RQ5 Can a multi-stage framework with multiple thresholding strategies achieve better robustness than single-method approaches under varying degradation conditions?

Key findings

• The proposed method achieved the highest F-measure score of 87.23% on the DIBCO 2013 benchmark, outperforming Otsu (80.04%) and Sauvola (85.02%).
• On the H-DIBCO 2014 dataset, the method achieved an F-measure of 92.40%, ranking first and significantly outperforming Otsu (91.63%) and Sauvola (86.83%).
• In the H-DIBCO 2016 benchmark, the method achieved an F-measure of 85.08%, ranking first and surpassing Otsu (86.59%) in F-measure, though Otsu had a slightly higher F-measure in this case.
• The method demonstrated the lowest DRD (Distance to Reference Distance) values across all datasets, indicating superior segmentation accuracy and minimal distortion.
• The PSNR values were consistently higher than most competing methods, with 18.35 dB on DIBCO 2013 and 19.09 dB on H-DIBCO 2014, indicating better image quality preservation.
• Visual inspection of binarization results confirmed that the proposed method produced the cleanest separation between text and background, with minimal noise and better character shape preservation compared to Niblack, Otsu, and Nick’s methods.