Elgammal, Kang, and den Leeuw (AAAI-18) introduced a stroke-level pipeline for computational analysis of artist drawings: individual strokes are segmented from raster scans, described by hand-crafted shape/tone features and learned sequence features, and aggregated to attribute drawings (Picasso, Matisse, Schiele) and reject commissioned fakes. The authors acknowledged that the method's validity and generality were unestablished: there is no expert ground-truth annotation of individual strokes, so segmentation quality and its downstream effect on attribution were unmeasured; data were too scarce for multi-class attribution; the corpus excluded shaded, hatched, and water-colored drawings; robustness across digitization conditions was asserted by design rather than tested; and key pipeline parameters were empirically tuned to one collection. The open problem is whether stroke-level analysis of drawings can be placed on a validated footing: expert-annotated stroke ground truth, quantified propagation of segmentation error into attribution and fake-detection decisions, and demonstrated generalization across artists, periods, drawing techniques, and imaging conditions, while remaining robust to imitation. Later work closed part of the gap (fine-grained artist identification from hatching lines by the same group; whole-image convolutional-neural-network and transformer authentication of paintings and drawings that bypass stroke segmentation), but no work to date provides stroke-level ground truth for historical drawings or measures segmentation-error propagation, so the validation-and-generalization core remains open.

Background

Deciding who made a drawing — attribution — and whether a work is genuine — authentication — has traditionally relied on connoisseurship: expert visual judgment of an artist's characteristic hand. Since the 2000s, computer-vision researchers have sought to quantify such judgment, mainly for oil paintings (e.g., statistical brushstroke analysis of van Gogh). Drawings are a harder case: the image consists of sparse, often crossing strokes made in diverse media (pen, ink, pencil, charcoal, watercolor), digitizations vary widely in resolution and paper/background appearance, and authenticated material per artist is scarce.

The source paper, 'Picasso, Matisse, or a Fake?' (Elgammal, Kang, and den Leeuw, 2018), proposed analyzing drawings at the level of the individual stroke, inspired by a classical manual methodology that quantifies stroke characteristics to judge authorship. The pipeline: a raster drawing is thinned to a one-pixel skeleton; strokes are segmented by linking skeleton endpoints through junctions (points where strokes cross or touch), using a graph search whose cost combines curvature with a junction potential of scale σ\sigma that penalizes paths passing through junction locations; each segmented stroke is then described by (i) hand-crafted global features (length, shape, tone statistics, normalized to physical millimeter units or length ratios to be invariant to digitization resolution) and (ii) learned local features — patches sampled along the stroke skeleton and fed to a GRU (gated recurrent unit, a recurrent neural network suited to variable-length sequences, trained with truncated back-propagation through time using an unroll length τ\tau). Drawing-level decisions aggregate per-stroke votes by majority vote, posterior-probability averaging, kk-certain voting (only strokes whose posterior probability for an artist exceeds threshold kk cast votes), or certainty-weighted voting.

Experiments used 297 line drawings from 1910–1950 (Picasso 130, Matisse 77, Schiele 36, Modigliani 18, and 36 by other artists; over 80,000 strokes) plus 83 drawings commissioned from five artists to imitate Picasso, Matisse, and Schiele. Results: 70–90% stroke-level classification accuracy, above 80% at the drawing level, and up to 100% fake-detection with hand-crafted features in the across-technique setting. Two failure modes were acknowledged: the GRU failed to detect fakes in the technique-specific setting (12.5% on Picasso fakes, 45% on Schiele fakes), and five-class attribution was ruled out as impractical given the achievable dataset sizes (49–75% test accuracy).

The paper also documented that segmentation could not be properly validated: no ground-truth stroke annotation of drawings exists. A small validation study — 14 drawings assessed by two artist annotators — found an average marked segmentation error of 12.94% and large disagreement between annotators (mean deviation 24.93, standard deviation 12%), highlighting both the difficulty and the ambiguity of stroke-level ground truth itself. The corpus was restricted to line drawings, explicitly excluding heavy shading, hatching (parallel lines used to convey tonality and volume), and water-colored strokes; resolution and paper/background invariance were built into the feature design but not systematically tested; and the parameters σ\sigma, τ\tau, patch radius (fixed vs adaptive to stroke width), and aggregation threshold kk were empirically selected on this one collection.

Later literature shifted much of the field toward whole-image or patch-level deep models: convolutional and transformer networks now attribute and authenticate paintings and drawings directly from pixels, and brushstroke- and surface-based authentication of oil paintings has advanced through patch analysis based on generative adversarial networks (GANs, in which a generator network learns to synthesize realistic images that fool a discriminator network), craquelure (crack-pattern) morphology, and fractal surface topography. These approaches sidestep stroke segmentation rather than validate it, so the specific question of whether stroke-level analysis — the level at which connoisseurs claim an artist's unconscious habits live — can be validated and generalized remains distinct from these advances.

Problem Statement

Can stroke-level computational analysis of digitized artist drawings be validated and generalized enough to support reliable attribution and authentication? Concretely, building on the pipeline and findings of Elgammal, Kang and den Leeuw (2018), a validated footing for stroke-level analysis encompasses: (1) ground truth — an expert-annotated corpus of individual strokes in digitized historical drawings at a scale sufficient to evaluate segmentation, with an annotation protocol that confronts the documented ambiguity of stroke boundaries (inter-annotator agreement measured and reported, not assumed); (2) error propagation — quantification of how stroke-segmentation errors of graded severity affect downstream stroke classification, drawing-level attribution, and fake detection, so that the method's claimed robustness is measured rather than asserted; (3) generality — demonstration, or rigorous refutation, of the claim that stroke-level features support attribution and imitation-resistant fake detection beyond pure line drawings of a few early-20th-century artists, i.e., across additional artists and periods and across drawing techniques including shaded, hatched, or water-colored marks, and across realistically varying digitization resolutions, papers, and backgrounds; (4) calibration — characterization of the sensitivity of the analysis to its principal design choices (in the source pipeline: junction scale, stroke-sequence modeling and unrolling, patch sizing, aggregation thresholds; in other stroke-level methods: their corresponding design choices) rather than tuning them to one collection. The question is about the stroke level of analysis in general, not about one fixed feature set or classifier: any method that segments and reasons about individual strokes falls within its scope. Together these components define what a validated footing requires — a full resolution addresses all four — but each is an independently refutable sub-result: settling any single component is a partial resolution, and a decisive refutation of one is itself a complete answer to that component. Because "enough" and "reliable" are not fixed a priori, the question turns on the concrete reliability criterion at stake — the stroke- and drawing-level accuracy, or error tolerance, at which attribution and authentication count as reliable.

References

  1. Elgammal, Ahmed; Kang, Yan; Den Leeuw, Milko (2018). Picasso, Matisse, or a Fake? Automated Analysis of Drawings at the Stroke Level for Attribution and Authentication. Proceedings of the AAAI Conference on Artificial Intelligence 32(1). DOI: 10.1609/aaai.v32i1.11313 (arXiv:1711.03536). https://doi.org/10.1609/aaai.v32i1.11313
  2. Ziaee, Shahrzad; Elgammal, Ahmed; Mazzone, Marian (2025). A Fine-Grained Artist Identification Method for Authentication and Attribution of Drawings Using Hatching Lines. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). DOI: 10.1109/CVPRW67362.2025.00203. https://doi.org/10.1109/CVPRW67362.2025.00203
  3. Frank, Steven; Frank, Andrea (2022). Analysis, Attribution, and Authentication of Drawings with Convolutional Neural Networks. International Journal of Arts and Technology 14(3). DOI: 10.1504/ijart.2022.10050958. https://doi.org/10.1504/ijart.2022.10050958
  4. Schaerf, Ludovica; Popovici, Carina; Postma, Eric (2023). Art Authentication with Vision Transformers. arXiv:2307.03039. https://arxiv.org/abs/2307.03039
  5. Boccuzzo, Sandro; Meyer, Deborah Desirée; Schaerf, Ludovica (2024). Art Forgery Detection using Kolmogorov Arnold and Convolutional Neural Networks. arXiv:2410.04866. https://arxiv.org/abs/2410.04866
  6. Azimi, Elhamsadat; Ashtari, Amirsaman; Ahn, Jaehong (2024). Patch-Based Oil Painting Forgery Detection Based on Brushstroke Analysis Using Generative Adversarial Networks and Depth Visualization. Applied Sciences 15(1):75. DOI: 10.3390/app15010075. https://doi.org/10.3390/app15010075
  7. Kim, Soojung; Park, Sang Min; Bak, Seongjin; Kim, Gyeong Hun; Kim, Chang-Seok; Jun, Joonja; Kim, Chang Eun; Kim, Kyujung (2022). Investigation of craquelure patterns in oil paintings using precise 3D morphological analysis for art authentication. PLOS ONE 17(7):e0272078. DOI: 10.1371/journal.pone.0272078. https://doi.org/10.1371/journal.pone.0272078
  8. Berkmans, Francois; Nys, Ludovic; Bigerelle, Maxence (2026). Preserving Van Gogh's painterly heritage: topographical and fractal insights in authentication. Surface Topography: Metrology and Properties. DOI: 10.1088/2051-672X/ae7136. https://doi.org/10.1088/2051-672X/ae7136
  9. Chen, Xiaojiao; Chen, Yonghao; Tang, Tan; Ge, Xin; Wang, Ruihan; Wang, Yifan; Wang, Xiaosong (2025). Multi-Modal Visual Analytic Approach for Attributing and Authenticating Ancient Chinese Paintings. Journal of Computer-Aided Design & Computer Graphics. DOI: 10.3724/sp.j.1089.2023-00350. https://doi.org/10.3724/sp.j.1089.2023-00350
  10. Kim, Nayoung; Jiang, Nan; Sun, Bangjie; Shin, Jaewon; Kim, Sojeong; Han, Jun (2026). IDraw: Artist Verification from Digital Drawing Images. arXiv:2608.01737. https://arxiv.org/abs/2608.01737
  11. Chen, Eric; Alves-Oliveira, Patricia (2026). Patch-Based Spatial Authorship Attribution in Human-Robot Collaborative Paintings. arXiv:2602.17030. https://arxiv.org/abs/2602.17030
  12. Castellano, Giovanna; Vessio, Gennaro (2021). Deep learning approaches to pattern extraction and recognition in paintings and drawings: an overview. Neural Computing and Applications 33(19):12263–12282. DOI: 10.1007/s00521-021-05893-z. https://doi.org/10.1007/s00521-021-05893-z
  13. Stork, David G. (2024). Computer Vision, ML, and AI in the Study of Fine Art. Communications of the ACM 67(5). DOI: 10.1145/3633454. https://doi.org/10.1145/3633454
  14. Jung, Soonchul; Kim, Jae Woo; Choi, Yoon-Seok; Jeon, Hyeong-Ju; Kim, Jin-Seo (2020). Generative Adversarial Network for Identifying Authors of Traced Images. 2020 International Conference on Information and Communication Technology Convergence (ICTC). DOI: 10.1109/ICTC49870.2020.9289592. https://doi.org/10.1109/ICTC49870.2020.9289592