PROTECT YOUR DNA WITH QUANTUM TECHNOLOGY
Orgo-Life the new way to the future Advertising by AdpathwayComputer vision researchers have developed DocRestormer, a text-structure-aware neural network designed to restore degraded historical document images and substantially improve the accuracy of downstream optical character recognition systems. Ancient manuscripts frequently endure physical decay, environmental exposure, and digitization artifacts that generate compound degradations such as ink fading, stroke breakage, edge erosion, and focal or motion blur. Conventional image restoration architectures often focus on pixel-level numerical approximations or global visual smoothing, which can obscure critical character topology and inter-stroke junctions. To overcome these constraints, the new model introduces explicit geometric mechanisms, anisotropic receptive fields, and guided structural priors that treat the recovery of text strokes as an integrated, structure-preserving process rather than a generic denoising task.
The study was designed around the distinct morphological challenges presented by historical Tibetan woodblock prints and manuscripts. Tibetan script exhibits strong horizontal line continuity combined with intricate vertical stacking of base letters, subjoined characters, and vowel signs within individual graphemic units. To provide a rigorous supervised benchmark for historical document restoration, the researchers compiled two paired datasets: the Lijiang Tibetan ancient document dataset, comprising two hundred image scrolls cropped into 2,348 training patches of 256 by 256 pixels, and the Beijing Tibetan ancient document dataset, consisting of 185 long-scroll images yielding 4,875 training patches. The original scrolls were strictly split into training, validation, and testing subsets prior to patch extraction, ensuring that patches derived from the same manuscript remained confined to a single partition to prevent data leakage.
Because authentic paired samples of pristine and deteriorated historical documents are exceptionally scarce, the research team constructed realistic degradation models tailored to the distinct baseline properties of each collection. For the relatively clear Lijiang collection, synthetic degradations simulated physical acquisition disturbances, including disk-shaped defocus blur modeled through circular point spread functions and irregular motion blur generated along randomized two-dimensional camera-shake trajectories. For the Beijing collection, which naturally exhibited mild ink fading, the authors formulated a structure-aware non-uniform degradation model. This method decomposed text regions using morphological dilation and erosion to isolate vulnerable stroke boundaries from stable interiors, applying bicubic upsampled noise fields and local background estimates to simulate realistic edge erosion, ink attenuation, and stroke discontinuities without generating artificial, homogeneous boundaries.
At the architectural core of DocRestormer is an adaptation of Restormer, a high-resolution vision Transformer that replaces quadratic spatial self-attention with transposed channel attention. To overcome Restormer’s rigid geometric inductive bias, DocRestormer incorporates a Deformable Convolution Local Enhancement module at the initial feature stage. Standard convolutions enforce fixed rectangular sampling grids that often blend degraded character strokes with adjacent paper background noise, diluting initial feature representations. The deformable module predicts eighteen content-adaptive spatial offsets across a nine-point kernel, allowing sampling points to dynamically shift along the curved and irregular contours of damaged strokes. This adaptive sampling decouples stroke foreground from background artifacts early in the network hierarchy, preserving fine boundary definitions before deep representation learning begins.
Within the deep Transformer blocks, the researchers replaced symmetric operators with Direction-Aware Multi-Dconv Head Transposed Attention and Multi-Scale Gated-Dconv Feed-Forward Networks. The direction-aware attention module deploys parallel 1-by-5 horizontal and 5-by-1 vertical depth-wise convolutions before computing channel covariance. This configuration directly aligns the network’s receptive field with the macroscopic horizontal baseline of document lines and the microscopic vertical stacking of Tibetan graphemic stacks, capturing anisotropic context without incorporating distracting information from neighboring lines. Operating in tandem, the multi-scale feed-forward network integrates parallel 3-by-3 and 5-by-5 convolutions with an identity branch, providing dynamic feature routing that accommodates substantial font scale variations ranging from large heading characters to fine textual annotations.
To provide high-level structural guidance, DocRestormer attaches a lightweight convolutional Stroke-Structure-Aware Branch after the primary decoder. This branch extracts local stroke geometries through four residual blocks without batch normalization and predicts a normalized stroke prior probability map. Rather than requiring laborious manual labeling, training is supervised through dynamic pseudo-targets computed directly from ground-truth images using adaptive intensity thresholding and Sobel edge detection. A two-stage optimization strategy prevents early, inaccurate structural predictions from corrupting the restoration backbone: the backbone and stroke branch are first trained independently with separate optimizers, and only after initial stabilization are they coupled in an aggregated guidance stage, where the predicted prior modulates normalized features and gates attention projections via a dedicated adapter.
Quantitative evaluations demonstrated that DocRestormer consistently outperformed existing general-purpose and document-oriented restoration models across both pixel-level reconstruction and text recognition metrics. On the Lijiang test set, which evaluated recovery under severe blur, DocRestormer achieved a peak signal-to-noise ratio of 24.364 decibels. When evaluated with the specialized Composite-CTC Tibetan recognizer, images restored by DocRestormer achieved a character error rate of 5.93 percent and a syllable error rate of 14.46 percent. These recognition figures represent absolute reductions of 1.25 and 2.64 percentage points, respectively, compared to images restored by the baseline Restormer architecture, which produced a character error rate of 7.18 percent and a syllable error rate of 17.10 percent from severely degraded input images that exhibited an unassisted character error rate of 54.60 percent.
On the Beijing dataset, which evaluated recovery under compound fading, reduced contrast, and moderate blur, DocRestormer achieved the leading peak signal-to-noise ratio of 35.952 decibels and tied for the highest structural similarity index of 0.921. Downstream recognition using Composite-CTC on these restored images achieved a character error rate of 3.26 percent and a syllable error rate of 7.98 percent, compared to 3.78 percent and 8.92 percent for baseline Restormer. An analysis of recognition error categories revealed that reductions in character error rates were driven primarily by a decrease in character substitution errors, dropping from 5,667 on the degraded Lijiang inputs to 989 under DocRestormer. This substantial reduction indicates that preserving local stroke connectivity and internal character geometry directly prevents recognizers from confusing degraded characters with morphologically similar counterparts.
To verify that the observed improvements were not artifacts of a specific downstream OCR design, the investigators conducted cross-recognizer validation across three distinct neural architectures: DAN, DAN+, and VAN. Trained on identical ground-truth splits, all three recognizers produced their lowest character and syllable error rates on images restored by DocRestormer, confirming that the structural enhancements generalized across different sequence-decoding frameworks. Stroke-level quantitative assessments using boundary F1-score and centerline Dice further revealed that DocRestormer significantly improved boundary completeness and topological continuity across all benchmarks, explaining why modest increases in global image metrics such as peak signal-to-noise ratio corresponded to disproportionately large gains in text recognizability.
Additional experiments confirmed that the DocRestormer architecture could be effectively trained on diverse document domains beyond Tibetan manuscripts. When retrained on the public Document Deblurring Dataset of English printed text and mathematical formulas, the model attained a peak signal-to-noise ratio of 28.459 decibels and a structural similarity index of 0.971, while on the ShabbyPages benchmark for document denoising and binarization, it achieved leading full-reference and perceptual fidelity scores. Ablation studies and capacity-controlled comparisons against a parameter-matched baseline confirmed that these performance gains stemmed from the task-oriented structural modules rather than increased model parameters, although qualitative error analyses indicated that minute, severely degraded features such as syllable delimiters remain susceptible to noise confusion and continue to present challenges for automated restoration.
Subject of Research: Document image restoration and historical manuscript recognition using structure-aware neural networks
Article Title: DocRestormer: a text-structure-aware network for document image restoration
Article References: Zong, X., Wang, W., Bi, S., Xiao, K., Li, Q., Mao, L., & Shi, X. (2026). DocRestormer: a text-structure-aware network for document image restoration. Journal of King Saud University Computer and Information Sciences, 38(9), Article 987. https://doi.org/10.1007/s44443-026-01369-z
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01369-z
Keywords: DocRestormer, document image restoration, optical character recognition, Tibetan ancient documents, Restormer, deformable convolution, stroke structure, image deblurring, ink fading, text degradation, character error rate, syllable error rate, vision transformer, feature modulation, historical manuscripts, document enhancement


2 hours ago
7




















English (US) ·
French (CA) ·