Art authentication remains a complex, multidisciplinary challenge that requires the integration of historical expertise, scientific analysis, and, increasingly, computational techniques (King, 2024). In recent years, deep learning has emerged as a powerful tool for image-based analysis, offering scalable, objective approaches that complement traditional expert judgment. Convolutional neural networks (CNNs) and other advanced architectures have demonstrated high levels of accuracy in art authentication by learning complex visual patterns directly from images.

However, despite the diversity of architectures and methodologies employed in prior work (see Section 2), deep-learning approaches share a fundamental requirement: access to large, labeled datasets for model training. Previous studies rely on hundreds or thousands of high-resolution images of authenticated artworks to enable neural networks to learn meaningful stylistic and compositional patterns. This dependence on extensive training data remains a key limitation, particularly when authenticating a single artwork, where the training set of verified examples may consist of only a handful of paintings. This data-scarce scenario renders standard deep, pretrained models, which contain millions of parameters, highly prone to overfitting (Bejani and Ghatee, 2021). Moreover, transfer learning approaches often rely on features learned from natural images (e.g., ImageNet), which emphasize high-level semantic content over low-level microtextural nuances required to distinguish a master from a workshop. A significant gap therefore exists for tailored methodologies capable of operating effectively under these severe data constraints, with emphasis on execution technique rather than semantic composition.

To address this gap, an approach based on a custom, shallow CNN and sliding-window-based feature extraction is introduced and applied to the authentication of a single painting with a disputed attribution. This patch-based approach mitigates data scarcity by substantially expanding the training set, shifting the model’s focus from global composition to learning localized, microstylistic features such as brushwork and texture (Sabha et al., 2024). Furthermore, the proposed model integrates multichannel inputs (RGB, grayscale, and edge maps) providing complementary visual cues; this technique has been shown to enhance the capture of subtle stylistic features (Ugail et al., 2023). The artwork, a Holy Family (see Figure 1A), has been linked to Paolo Veronese, although its authorship remains uncertain, and may originate from the master, Veronese’s workshop, or his disciples (Blanc et al., 2023; López-Baldomero et al., 2023). Training on a restricted dataset of Holy Family paintings attributed to Veronese and his circle, is intended to identify stylistic features that may clarify the painting’s origin, offering a reproducible and data-driven complement to expert judgment.

The main contributions of this study are as follows:

The remainder of this paper is organized as follows: Section 2 reviews related work; Section 3 describes the dataset and the proposed methodology; Section 4 presents the experimental results; and Section 5 discusses the results and presents the conclusions.

Deep learning has emerged as a powerful approach for image-based analysis in the cultural-heritage domain. CNNs and advanced architectures, such as vision transformers, have demonstrated high levels of accuracy in art authentication by learning complex visual patterns directly from images (Elgammal et al., 2018; Dobbs et al., 2023; Ugail et al., 2023; Chen et al., 2024; Schaerf et al., 2024).

Specifically, in Schaerf et al. (2024), the use of vision transformers for the authentication of Van Gogh paintings was explored, showing that deep-learning models can outperform handcrafted-feature approaches in both precision and interpretability. Similarly, Dobbs applied large-scale CNN-based classification framework to contemporary artworks, achieving more than 91% accuracy and highlighting deep learning’s potential for scalable authentication across diverse artistic styles (Dobbs et al., 2023).

In Ugail et al. (2023), the effectiveness of transfer learning in attributing Renaissance paintings, specifically works by Raphael, using pretrained CNNs fine-tuned on curated datasets was demonstrated. Their approach provided valuable insights into stylistic attribution and supported expert judgment in complex cases. The study in Chen et al. (2024) further reinforced the utility of CNNs in fine-art recognition, emphasizing their role in extracting stylistic and compositional features for authentication tasks. Earlier work in Elgammal et al. (2018) introduced a deep-learning framework for analyzing stylistic evolution in art history, laying the groundwork for subsequent applications in authentication and attribution.

Beyond the cultural-heritage domain, the efficacy of shallow and custom-designed CNNs has been demonstrated in other fields requiring high levels of precision under specific constraints. Recent advances show that compact models, when properly optimized, can rival deeper architectures. For instance, Radojcic proposed a two-layer TinyML approach for plant disease classification (Radojcic et al., 2026), highlighting the computational efficiency of shallow networks. Similarly, the studies in Zivkovic et al. (2025) and Basha et al. (2021) employed metaheuristic optimization strategies to enhance CNN design for ocular-disease diagnosis and general-purpose applications, respectively. These studies reinforce the premise that tailored, shallow architectures constitute a robust solution for specialized tasks where massive datasets for pretraining are unavailable or unsuitable.

This section presents the experimental results, first detailing the CNN’s cross-validation performance (Section 4.1), then the authentication results for the test paintings and the painting under investigation (Section 4.2), and finally the comparison with the baseline model MobileNetV2 (Section 4.3).

This study addressed the challenge of authenticating paintings under conditions of data scarcity by developing a patch-based CNN pipeline tailored to stylistic analysis. The methodology combined sliding window feature extraction with multichannel inputs (RGB, grayscale, and edge maps) to capture complementary visual cues and implemented painting-level cross-validation to prevent data leakage. By expanding the dataset with localized patches and integrating regularization strategies, the approach aimed to enhance generalization while preserving stylistic fidelity. The CNN trained with these patches distinguished authentic Veronese works from non-Veronese paintings at both painting- and patch-levels. Authentic Veronese paintings show high painting-level probabilities (ranging from 80.3 to 99.9%) and Veronese probability heatmaps with broad, contiguous high-probability regions (see Figure 7), whereas non-Veronese paintings exhibit lower painting-level scores (ranging from 11.9 to 36.7%) and fragmented, localized regions of high probability (see Figure 6). The painting under authentication yields an intermediate painting-level score (61.0%) with extensive high-probability zones over stylistically salient passages, interspersed with some lower-probability areas (see Figure 8).

The sliding window design (64 × 64 patches, 65% overlap) expanded the training data from a small number of paintings to thousands of localized samples, encouraging the CNN to learn microtextural and brushwork signatures while controlling overfitting via dropout, L2 regularization (weight decay), batch normalization, the Adam optimizer, and early stopping. The high overlap mitigates boundary artifacts and stabilizes local predictions, an effect analogous to the overlap-tile inference widely used in biomedical image analysis (Ronneberger et al., 2015). The multichannel design (RGB, grayscale, and edge map) leveraged complementary cues: chromatic information, luminance-texture structure, and gradient-defined contours. The Canny edge detector is well-suited here due to its double-threshold mechanism and robustness to noise, which supports capturing subtle, style-relevant edge patterns (Canny, 1986).

The 6-fold painting-level cross-validation was essential to prevent data leakage from overlapping patches, an issue known to inflate performance in tile-based pipelines and subject-repeated imaging datasets (Bussola et al., 2019; Rumala, 2023). Cross-validation confirmed strong overall performance (see Table 4; Figure 5): accuracy 94.51%; precision 94.75%; F1 score 95.94%; sensitivity 97.15%; specificity 89.23%; G-mean 93.11%; Cohen’s kappa 0.87; and AUC-ROC 0.99. This set of metrics provides a balanced view: AUC-ROC summarizes rank discrimination (Fawcett, 2006); G-mean emphasizes balanced sensitivity and specificity under class imbalance (Kuncheva et al., 2019); and Cohen’s kappa quantifies agreement beyond chance (Warrens, 2014). The validation metrics confirmed the CNN model’s high accuracy and exceptional discriminative power. The model was particularly effective at identifying authentic Veronese paintings (sensitivity = 97.15%), with slightly lower performance for non-Veronese paintings (specificity = 89.23%).

The heatmap visualization of Veronese probability provides critical insight into CNN’s decision-making process. In the painting under authentication (see Figure 8), high-probability regions cluster around stylistically salient elements such as anatomical contours and tonal transitions, whereas low-probability regions dominate less textured or peripheral areas. This uneven distribution suggests that the model detects Veronese-like features selectively rather than uniformly across the composition, helping to explain the intermediate global score of 61%. From a methodological perspective, this outcome reflects the patch-based classification strategy: aggregate probabilities capture local stylistic heterogeneity rather than enforce global uniformity. From an interpretive standpoint, the observed pattern may indicate (1) an authentic work with workshop participation; (2) an authentic work with later interventions or restoration that may alter original textures; or (3) a high-quality workshop production under close supervision. Consequently, while the CNN identifies meaningful stylistic affinities, the evidence remains insufficient to support definitive attribution.

The comparison between the painting under authentication and the non-Veronese replicas of the composition (paintings 7, 8, and 9) highlights the CNN’s capacity to assess stylistic consistency at a global level while acknowledging local similarities. Although the non-Veronese paintings exhibit localized high-probability regions (see Figure 6), primarily around anatomical details and drapery folds, their overall scores remain low (ranging from 32 to 37%), indicating limited stylistic coherence. In contrast, the painting under authentication shows broader regions of high Veronese probability and a substantially higher global score (61%), suggesting a more pervasive presence of Veronese-like features. This pattern supports the hypothesis of partial authenticity or strong stylistic influence for the painting under authentication and confirms that the CNN does not rely solely on compositional similarity but captures nuanced textural and structural cues. The residual probabilities in non-Veronese paintings may arise from shared iconographic elements, underscoring the importance of interpreting probability heatmaps in conjunction with global metrics rather than in isolation. A key detail is that in all non-Veronese paintings, the area corresponding to St. Joseph in the upper-right corner consistently yields very low probabilities, whereas in the painting under authentication the corresponding region exhibits high probabilities (see Figure 8). This contrast indicates that the CNN captures subtle stylistic cues in localized passages that are absent or less pronounced in non-Veronese paintings despite compositional similarity.

The probability heatmaps for authentic Veronese paintings (see Figure 7) reveal broad, continuous regions of high probability. This spatially cohesive distribution contrasts sharply with the fragmented patterns observed in non-Veronese works and supports the CNN’s capacity to capture stylistic coherence that extends beyond isolated details. The consistently high painting-level scores for Veronese paintings (ranging from 80.3 to 99.9%) indicate that the model generalizes effectively to unseen authentic works. These findings support the patch-based approach as a robust strategy for identifying nuanced textural and structural cues characteristic of Veronese’s technique.

The comparison between paintings 4 (authentic Veronese) and 10 (copy) is particularly noteworthy. Despite compositional similarity, the authentic work yields a high painting-level probability (89.2%) and extensive high-probability regions (Figure 7). In contrast, the copy yields a low painting-level probability (11.9%) with sparse and fragmented high-probability regions (Figure 6). This supports the CNN’s capacity to differentiate stylistic coherence from superficial compositional similarity.

Notably, one authentic Veronese painting (see painting 2 in Figure 7) exhibits a sharply defined low-probability region confined to a rectangular area in the upper-left corner. Two nonexclusive explanations are plausible: (1) the model may struggle to classify uniform, low-texture passages that provide weak class-specific signals; (2) such areas may indicate workshop participation in secondary passages, an documented practice in 16th-century Venice in which masters concentrated on principal figures while assistants executed peripheral and ornamental elements (Gisolfi, 2017).

To comprehensively benchmark the proposed architecture, a comparative experiment with MobileNetV2 (Sandler et al., 2018) via transfer learning using a 224 × 224 × 3 input was conducted. The results indicated critical limitations in applying standard deep learning models to this particular domain. Although MobileNetV2 achieved high metrics during cross-validation (accuracy: 97.23%), it failed to establish a discriminative boundary in the test dataset, yielding overlapping probability ranges to Veronese works (62.8–73.0%) and non-Veronese paintings (up to 72.7%). Two technical factors account for this generalization failure when compared with the proposed architecture. First, the loss of resolution: the required upsampling from 64 × 64 to 224 × 224 pixels effectively acts as a low-pass filter, smoothing out high-frequency microtextures essential for distinguishing Veronese brushwork. Consequently, the pretrained model likely overfitted to macroscopic features, such as color palettes and semantic content and composition, shared by both originals and workshop productions. Second, input rigidity: by restricting the input to RGB channels, MobileNetV2 excluded the explicit grayscale intensity and edge maps used in the proposed pipeline, thereby losing critical topological information about brushstroke dynamics that proved decisive in the shallow CNN.

Notably, for the Holy Family painting under authentication, MobileNetV2 predicted a Veronese probability of 63.6%, closely matching the 61.0% obtained by the proposed shallow CNN. This convergence suggests that the classification of the artwork as an intermediate or workshop production is a robust signal, persisting across different model architectures. Although MobileNetV2 lacks sufficient specificity to reject clearly non-Veronese works, its agreement with the proposed specialized shallow CNN for the disputed artwork supports the reliability of the assessment: the painting exhibits a hybrid visual structure that is neither accepted as a fully authentic Veronese work nor rejected as a mere Veronese-style replica.