Transformers Learn to Recognize You by Your Handwriting Alone

Every piece of handwriting carries a signature that has nothing to do with the words being written. The slant of a loop, the pressure implied by a stroke’s thickness, the peculiar way a writer connects letters or leaves them stranded — these habits are so distinctive that forensic examiners have relied on them for more than a century. Now a team of researchers at Kim Il Sung University in Pyongyang has built an artificial intelligence system that learns to read those habits automatically, and their results suggest that transformer networks, the same architecture behind modern language models, can be retooled to identify writers from scanned pages of text without ever needing to know what the text says.

The new study, published in the journal Multimedia Tools and Applications, addresses a task known as text-independent offline writer identification. Offline means the system works from static images of handwriting — photographs or scans of paper documents — rather than from digital pen strokes captured in real time. Text-independent means the system cannot rely on recognizing specific words or characters: it must judge authorship from style alone, even when the documents being compared contain entirely different content. This is by far the harder version of the problem, and it is the version that matters most in practice, because forensic investigators, archivists, and historians rarely get to choose what text a writer happened to leave behind.

The difficulty stems from the sheer variability of human handwriting. The authors of the study point out that a person’s script shifts with age, education, emotional state, and even the intention behind the writing, while the appearance of a scanned document is further distorted by the writing instrument used and the conditions of acquisition. A ballpoint pen on glossy paper produces a different image than a fountain pen on rough fiber, even when the same hand guides both. Any system that hopes to identify writers reliably must therefore extract features that are stable across all of these nuisance factors while remaining sensitive to the deep, persistent individuality of motor habits.

The core of the proposed framework is a modified vision transformer built around what the researchers call local-global attention blocks. In a conventional vision transformer, an image is chopped into small patches, and a self-attention mechanism lets every patch exchange information with every other patch through learned projections known as queries, keys, and values. This global attention is powerful, but it has a weakness for handwriting: the most telling stylistic evidence often lives in fine local structures — the curve of a single stroke, the junction where two lines meet — and standard projection layers can blur or wash out those details. The team’s solution is to replace the standard query, key, and value projections with depth-wise separable convolutional projections, a lightweight convolutional operation that processes each channel of the patch representation independently before combining them. The effect is that the network preserves local handwriting structures while still allowing patches to communicate across the whole page.

The second innovation is a global average pooling branch woven into the architecture through a gated skip connection. Pooling across the entire feature map gives the network a compressed summary of the page’s overall writing style — spacing, rhythm, and layout tendencies that no single patch can convey. The gate, a learned switch, lets the model decide dynamically how much of this global context to blend into the local features at each block. The result is an architecture that simultaneously attends to the microscopic texture of strokes and the macroscopic character of the page, mirroring the way a human examiner might zoom in on a suspicious letterform and then step back to judge the overall flow of the script.

Getting the input right is half the battle, and the researchers designed a careful preprocessing pipeline to feed their network. Scanned pages are first segmented into individual text lines, and then words are recombined into coherent patch inputs. This matters because handwriting does not respect neat rectangular boundaries: ascenders and descenders from adjacent lines can overlap, and word spacing varies wildly between writers. By reconstructing word-level regions after line segmentation, the framework gives the transformer patches that correspond to meaningful units of writing rather than arbitrary crops, so the attention mechanism operates on fragments that actually carry stylistic information.

At the matching stage, the system condenses each document into a 512-dimensional feature representation — a numerical fingerprint of writing style — and performs page-level writer matching against a reference database. Page-level matching is a deliberate design choice. Earlier systems often compared documents line by line and then aggregated the results, but the authors show that their page-level strategy improves Top-1 accuracy, meaning the correct writer appears at the very top of the ranked candidate list more often. Intuitively, a full page offers a richer and more redundant sample of a writer’s habits than any single line, and the local-global attention architecture is precisely what allows the network to exploit that abundance without drowning in it.

The experimental evidence comes from three widely used benchmark datasets: IAM, CVL, and Firemaker. These collections span different languages, scripts, and acquisition conditions, which makes them a demanding test of generalization. IAM, drawn from English sentences copied by hundreds of writers, tests performance on unconstrained modern script. CVL concentrates on single-page samples from a large writer pool, stressing the system’s ability to discriminate when it has only limited material per person. Firemaker adds further variety in writing styles and document conditions. Across all three, the proposed framework achieved strong identification performance and consistently outperformed the line-level matching strategy, supporting the authors’ argument that page-level reasoning combined with local-global attention is the right recipe for the task.

The significance of this work extends beyond a leaderboard result. Writer identification sits at the intersection of biometrics, forensics, and cultural heritage. In forensic contexts, a robust text-independent system could help narrow suspect lists from questioned documents. In archives and libraries, it could assist in attributing unsigned manuscripts, detecting forgeries within historical collections, and tracing the hands behind medieval scripts — an application area where deep learning for writer identification has already shown promise. The architectural idea at the heart of the study, blending convolutional inductive biases into transformer attention, also reflects a broader trend in computer vision, echoing hybrid designs explored in models such as LeViT, Swin Transformer, and RegionViT, which similarly reconsider where pure global attention helps and where local structure must be protected.

There are, of course, limits to what the published results establish. The study reports performance on benchmark datasets rather than on casework material, and real forensic documents introduce degradations — folds, stains, partial pages, mixed writers — that benchmarks only approximate. Handwriting also changes over a person’s lifetime, and the authors themselves enumerate the subjective and objective factors that make the problem hard. Still, the framework’s design choices are well motivated and its evaluation spans three independent datasets, which lends credibility to the central claim: that depth-wise separable convolutional projections and gated global context give transformers the sensitivity handwriting demands. As handwriting analysis joins the growing list of domains being reshaped by attention-based models, this study offers a concrete demonstration that the architecture of the AI era can be tuned to one of humanity’s oldest biometric traits — the way each of us, uniquely, puts pen to paper.

Subject of Research: Text-independent offline writer identification using a transformer with local-global attention blocks

Article Title: Text-Independent Offline Writer Identification with Local–Global Transformer Attention

Article References: Ri, C.-Y., Choe, K.-H., Ri, M.-G., Choe, H.-S., & Ri, S.-J. (2026). Text-Independent Offline Writer Identification with Local–Global Transformer Attention. Multimedia Tools and Applications, 85(10), Article 797. https://doi.org/10.1007/s11042-026-21932-0

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21932-0

Keywords: writer identification, handwriting analysis, transformer, vision transformer, self-attention, biometrics, forensics, deep learning, computer vision, IAM dataset, pattern recognition, local-global attention