Enterprise AI Analysis: Document Image Binarization

SEADUNet: A Novel Architecture for Multilingual Binarization

SEADUNet introduces a novel binarization approach that combines multi-scale convolutional attention feature fusion (EMCAM) with spatial-channel reconstructed convolution (SCPConv) techniques. This architecture is designed to handle the complexities of multi-script ancient document images by enhancing feature mapping and focusing on prominent areas within the images.

• SCPConv Module: Replaces traditional convolution in the U-Net encoder to extract rich feature representations and reduce spatial/channel redundancy, crucial for processing blurred and ink-bleeded text.
• Spatial Group Enhancement (SGE): Dynamically adjusts sub-feature importance through location-specific attention, enabling autonomous enhancement and noise suppression.
• EMCAM Attention Mechanism: Integrated into the U-Net decoder, this module uses multi-scale deep convolutional blocks and includes a Channel Attention Block (CAB), Spatial Attention Block (SAB), and Efficient Multi-Scale Convolutional Block (MSCB) to refine feature mappings, enhancing context retention and feature fusion.

This integrated approach significantly improves the quality and accuracy of binarization, making it highly adaptable to diverse writing styles and degradation levels found in ancient documents.

Superior Performance Across Diverse Scripts and Degradations

Experiments were conducted on the newly established Multilingual Ancient Document Image Binarization Dataset (MADIBD2024-16), comprising 3,200 annotated image pairs from 16 distinct historical scripts. SEADUNet demonstrated impressive performance, achieving an F-Measure (FM) of 95.54%, a pseudo F-Measure (p-FM) of 95.98%, a Peak Signal to Noise Ratio (PSNR) of 20.67 dB, and a Distance Reciprocal Distortion (DRD) of 2.59.

Ablation studies confirmed the synergistic benefits of SCPConv, SGE, and EMCAM, showing that their combined use leads to the best performance. Compared to both traditional and cutting-edge deep learning methods, SEADUNet proved particularly adept at handling the binarization of multi-script ancient document images, showcasing robust noise reduction and character preservation capabilities. Additional validation on other ancient script datasets further substantiated the model's universality and practicality.

MADIBD2024-16: A New Benchmark for Ancient Document Research

To address the scarcity and suboptimal quality of existing datasets, this paper introduces the Multilingual Ancient Document Image Binarization Dataset (MADIBD2024-16). This rigorously curated collection includes 3,200 annotated image pairs spanning 16 distinct historical scripts, with an 8:2 training to test set ratio. The dataset is crucial for evaluating and advancing document binarization algorithms, offering a standardized benchmark.

Its significance lies in its comprehensive multi-language coverage, reflecting diverse preservation media (paper, bamboo, silk, cotton, wood) and presenting a rich array of challenges from varying noise types. This high-quality data foundation is essential for researchers to explore ancient books across linguistic and cultural contexts, facilitating the development of robust and accurate binarization technology for cultural heritage preservation.