Paper authors Hassan Ugail, Jan Ritch-Frel, Irina Matuzava, and David G. Stork have published their work in PLOS One Journal . You can read their article, “Verification of historical sketches via one-class learning on compact feature representations,” on the PLOS website .
The synopsis and introduction for the article listed below. Abstract Historical sketch authentication is challenging because securely attributed reference sets are often small, and stylistic evidence is carried primarily by line, texture, tonal variation, and mark-making. We present a reproducible framework for verifying historical sketches using artist-specific one-class autoencoders trained on compact handcrafted feature representations. Ten artist models were trained using authenticated sketches from six open-access cultural heritage collections. Each drawing was represented by five interpretable descriptors, namely, Fourier-domain energy, Shannon entropy, global contrast, Grey-Level Co-occurrence Matrix homogeneity, and box-counting fractal complexity. The system was evaluated using a biometric-style verification protocol in which each artist model was tested on genuine held-out works and impostor works by other artists. On the primary evaluation partition of 900 decisions, comprising 90 genuine and 810 impostor trials, the method achieved 87.6% balanced accuracy, 77.8% True Acceptance Rate, 2.6% False Acceptance Rate, 0.748 Matthews Correlation Coefficient, and 11.4% Equal Error Rate. Performance remained stable across 20 repeated random train/test splits. The proposed model also outperformed Gaussian and one-class SVM baselines, while pretrained ResNet50 and EfficientNet-V2 feature representations performed substantially worse in this data-scarce setting. Leave-one-feature-out ablation confirmed that all five descriptors contributed positively, with fractal complexity and GLCM homogeneity providing the strongest individual contributions. Error analysis revealed structured false-accept pathways to be consistent with stylistic proximity between artists. The framework provides transparent, reproducible, and interpretable quantitative evidence for historical sketch verification. It is intended to support, not replace, expert connoisseurship in attribution settings where available reference corpora are limited. Introduction 1.1. Context and motivation
The authentication and attribution of historical artworks are central concerns in art history, conservation, and the art market. For works on paper, these concerns are intensified by the material and documentary conditions under which drawings survive. For example, sketchbooks may be dispersed, sheets may be trimmed or mounted, and many works exist in multiple states or workshop contexts. Connoisseurship remains indispensable in this domain, yet it is intrinsically difficult to formalise, reproduce, and quantify, particularly when disputes arise, and decision-makers require transparent evidence beyond expert opinion [ 1 , 2 ]. Computer vision, machine learning, and artificial intelligence offer an additional, complementary approach to analysis [ 3 ]. In particular, the logical framework of biometric verification provides an appealing analogue. For example, a test sample is verified against a target identity, and system performance is characterised by false acceptance and false rejection under explicit operating points. In art authentication, the “identity” is the target artist, and impostor trials represent non-target artists (and, in principle, forgeries). This framing falls in line with the open-set nature of attribution, where it is rarely possible to enumerate all plausible non-target classes of potential authors of the artwork in question. Recent advances in machine learning have demonstrated promising results for image-only art attribution when large labelled datasets are available. A comprehensive review of the use of artificial intelligence in art authentication is provided by Cetinic and She (2022), who document the field’s shift from traditional computer vision methods towards deep neural networks [ 4 ]. Building on this trend, deep convolutional and attention-based architectures have been applied to artist attribution and style analysis across large corpora, achieving strong performance when abundant labelled data are available [ 5 ]. Despite these advances, such approaches generally depend on thousands of examples per artist, a requirement rarely met in the context of historical sketches or for most historical painters [ 3 ].
1.2. Data scarcity and the rationale for one-class verification
Deep supervised attribution methods generally require large labelled datasets and benefit from broad negative sampling [ 6 ]. Historical sketches rarely satisfy these conditions. Even for major artists, the number of authenticated drawings available as consistent digital surrogates is limited. Moreover, intra-artist variability can be substantial because sketches are often rapid, exploratory studies rather than finished works [ 5 ]. These constraints can be addressed through one-class learning, where the model learns a representation of the authentic distribution of a single artist and flags deviations as anomalous [ 7 , 8 ]. One-class verification is particularly appropriate when negative classes are heterogeneous, incompletely characterised, or strategically adversarial (as in forgery scenarios) [ 9 ]. This approach also fosters methodological unity and consistency, avoiding the methodological variations inherent in different analyses based on unequal choices of non-target training data. The challenge of limited training data is not unique to art authentication. Few-shot learning approaches have been explored in various domains [ 10 , 11 ], but these typically still require more examples than are available for many historical artists. Transfer learning from pre-trained models offers another avenue [ 12 ], but the domain gap between natural images and historical sketches can be substantial. One-class learning sidesteps these issues by focusing solely on modelling the authentic distribution without requiring comprehensive negative examples.
1.3. The role of handcrafted features in data-scarce settings
Whilst end-to-end deep learning has dominated recent work in computer vision, handcrafted features retain important advantages in data-scarce scenarios. They reduce sample complexity through dimensionality reduction, provide interpretability enabling expert validation, incorporate domain knowledge, and offer greater robustness to distribution shift than learnt representations when training sets are small [ 5 , 12 ]. For sketch authentication, colour information is limited, and style is expressed primarily through marks, shading, and tonal distribution. Carefully designed handcrafted features can capture essential stylistic signals whilst remaining trainable with minimal data. Texture features derived from Grey-Level Co-occurrence Matrices have proven effective for distinguishing artistic techniques [ 13 ]. Frequency domain analysis reveals characteristic rhythmic patterns in artists’ marks [ 14 ]. Fractal analysis captures the hierarchical complexity of mark-making [ 15 – 17 ]. Information-theoretic measures quantify tonal complexity and distributional properties [ 18 , 19 ]. By combining features that have been proven to be informative in many domains (including art analysis), we can construct compact yet informative representations suitable for one-class learning.
1.4. Scope and contributions
This study develops a reproducible verification framework for sketch authentication under severe corpus size constraints. The central methodological contribution is an artist-specific one-class autoencoder verifier trained on interpretable handcrafted features well suited to line-dominant media. The empirical contribution is a multi-artist evaluation across ten historical artists, reporting both pooled and artist-specific biometric metrics with Wilson confidence intervals, partition robustness evidence from repeated random sub-sampling, a controlled leave-one-feature-out ablation, and a structured attribution of false-accept pathways to identify systematic confusability between artists. Our principal contributions are as follows:
- A novel application of one-class autoencoder architecture to historical sketch authentication, demonstrating effective discrimination despite severe data scarcity (20 training images per artist).
- - Identification and formal definition of five literature-motivated handcrafted features—Fourier energy, Shannon entropy, contrast, GLCM homogeneity, and box-counting fractal dimension—selected to capture distinct properties of artistic style in line-dominant media, with feature necessity confirmed empirically by ablation.
- - Comprehensive multi-artist evaluation using a rigorous biometric verification framework with 900 trials, reporting all metrics with Wilson binomial confidence intervals appropriate for small sample sizes, and using MCC and balanced accuracy as primary discrimination summaries to account for class imbalance.
- - Partition robustness analysis via 20 independent repeated random sub-samplings (seeds 0–19) of the 29-image corpus per artist, establishing that the primary reported results are representative rather than an artefact of a single train/test split.
- - A systematic leave-one-feature-out ablation study using a fixed-capacity architecture across all conditions to isolate feature contribution from model-capacity effects, identifying fractal dimension and GLCM homogeneity as the most informative individual features.
- - Pairwise confusion attribution revealing structured error relations consistent with art-historically interpretable stylistic proximity.
- Article Source: Verification of historical sketches via one-class learning on compact feature representations Ugail H, Ritch-Frel J, Matuzava I, Stork DG (2026) Verification of historical sketches via one-class learning on compact feature representations. PLOS ONE 21(6 e0344796. https://doi.org/10.1371/journal.pone.0344796 Photo Credit: Anton van den Wyngaerde via Wikimedia Commons The post Visual Computing Analysis PLOS One Paper to Verifying Artist Drawings and Sketches appeared first on Independent Media Institute .