Artistic style classification sits at the intersection of computer vision and digital art history. A style label encapsulates the shared techniques, philosophies, and visual idioms of a movement, and automated classification supports applications ranging from museum curation and art education to large-scale analysis of cultural heritage collections. The task is genuinely difficult: paintings within the same style can vary enormously, while stylistically distinct works can share overlapping features. Unlike ordinary image recognition, which often leans on object shapes or semantic content, style recognition depends on abstract and perceptual attributes—brushstroke direction, density, and smoothness; texture roughness and layering; color palettes and contrasts; and the spatial arrangement of elements on the canvas.

Convolutional neural networks, including architectures such as VGG16, VGG19, Inception, EfficientNet, and ResNet, have been the dominant paradigm for this image-recognition problem because their inductive biases favor local feature extraction. That strength, however, may become a bottleneck. The defining elements of a style—the dynamic composition of a Baroque piece, the blended skies of Impressionism, or the fragmented geometry of Cubism—require a model that integrates information across the whole image from the outset. Vision Transformers address this by treating an image as a sequence of patches and applying a self-attention mechanism that captures global dependencies between all patches simultaneously, regardless of their spatial proximity, making the architecture well suited to the macro-scale analysis that art style recognition demands.

Transformers, however, are notoriously data-hungry, and heritage datasets are typically small because artworks are fixed and irreplaceable. The researchers tackled this constraint with a structured training pipeline combining transfer learning, data augmentation, the AdamW optimizer, and a cosine annealing learning-rate scheduler. The backbone is a lightweight ViT-Tiny/16 configuration: input images are resized to 224 by 224 pixels and divided into non-overlapping 16 by 16 patches, yielding 196 patches arranged in a 14 by 14 grid. Each patch is flattened and linearly projected into an embedding vector of dimensionality 192, and a learnable classification token is prepended to the sequence, with learnable positional embeddings added to preserve spatial ordering that the Transformer would otherwise not know.

The Transformer core consists of 12 sequential encoder blocks, each containing a multi-head self-attention module with three attention heads, normalization layers, residual connections, and a feed-forward network using the GELU activation function. In self-attention, the input token embeddings are transformed into Query, Key, and Value matrices; the query of each token is compared against the keys of all others, and feature information is aggregated according to the resulting attention weights. With three heads, the model can represent multiple types of relationships among patches in parallel. Unlike a CNN, whose receptive field is fixed by kernel size, self-attention dynamically captures features across a much wider range. The final class-token representation, which aggregates global information from all patches, feeds the five-class classification head.

The model was evaluated on the Persian Miniatures dataset, a curated collection of 890 images drawn from the Islamic Painted Page archive and refined through manual cleaning, cropping, and removal of marginalia. The images span five historical and stylistic schools: Herat with 202 images, Qajar with 150, Shiraz-e Avval with 193, Tabriz-e Avval with 190, and Tabriz-e Dovvom with 155. The authors chose this dataset for its diverse class labels, high-quality curation, and prior validation in computational research. Because the dataset is small, the team used five-fold cross-validation, applying augmentation—horizontal flips with probability 0.5, rotations between minus 15 and plus 15 degrees, 10-degree shearing, and brightness, contrast, and saturation adjustments of 0.1—only to the training folds.

Training ran for 20 predefined epochs per fold with a batch size of eight, an initial learning rate of 0.001, and a weight decay of 1 times 10 to the minus 4. The backbone was instantiated from the timm library with pretrained weights and fine-tuned end-to-end, meaning all trainable parameters, including the encoder layers and the new classification head, were updated rather than frozen. The resulting model has approximately 5.53 million trainable parameters, substantially fewer than larger ViT variants—a deliberate choice to reduce computational cost and overfitting risk on the limited dataset. Notably, the authors report that their original environment did not record the exact timm package version or checkpoint metadata, so they disclose only the model identifier and initialization flag rather than asserting an unverified pretraining corpus.

Across the five evaluation folds, AOViT achieved a mean accuracy of 0.9385 with a standard deviation of 0.0079, precision of 0.9394, recall of 0.9385, and an F1-score of 0.9384, with approximate 95 percent confidence intervals spanning roughly 0.929 to 0.948 for the main metrics. An ablation study revealed a nuance worth attention: data augmentation alone actually lowered accuracy, from 91.73 percent in the baseline ViT to 90.73 percent, and the learning-rate scheduler configuration also fell slightly below baseline at 91.62 percent. The largest individual improvement came from AdamW, which decouples weight decay from the gradient-based parameter update and lifted accuracy to 93.18 percent. Combining all components produced the best result of 93.85 percent, though the authors caution that augmentation’s independent or synergistic contribution cannot be definitively determined because not all interactions were separately evaluated.

Descriptive comparisons under identical preprocessing and training settings showed AOViT outperforming several CNN architectures, including InceptionV3, VGG-19, VGG-16, EfficientNet-B4, and Xception, as well as three hybrid Transformer-CNN models built on InceptionV3, ResNet50, and EfficientNet-B4 backbones. The margins over the hybrids were 2.85, 7.18, and 11.63 percentage points respectively, and AOViT trained faster than every hybrid, taking 659.294 seconds compared with 957.721 seconds for ViT plus InceptionV3. The authors are careful to state that these numerical differences are descriptive only; without sample-level paired predictions, no McNemar’s test or paired-bootstrap significance results are claimed, and comparisons with previously published studies on different datasets and protocols are likewise not evidence of statistically significant superiority.

To open the black box, the team applied Grad-CAM, an explainable artificial intelligence technique that generates heatmaps highlighting the image regions contributing most strongly to a prediction. For an illustrated example, the strongest activations fell on the central seated figure and a nearby vessel, with moderate activation over surrounding foliage and low attribution for the standing figure, decorative frame, and corners; thresholding at 0.7 isolated only the highest-attribution regions. Quantitatively, AOViT achieved the lowest Deletion AUC of 0.118 and the highest Insertion AUC of 0.642 among the evaluated models, indicating stronger explanation faithfulness than the baseline ViT, VGG-16, or VGG-19 under that protocol. The authors stress these metrics measure faithfulness, not spatial localization accuracy, since the dataset lacks ground-truth region annotations. They also acknowledge the study’s limits: the dataset is small and domain-specific, cross-dataset validation was not performed, and Transformer computational demands may constrain real-time heritage applications. Future work, they suggest, should test larger and more diverse collections, explore cross-dataset generalization, and investigate lightweight or hierarchical Transformer architectures and efficient CNN-Transformer hybrids.

Subject of Research: Artificial Intelligence

Article Title: AOViT: an adaptive optimized vision transformer model for the classification of heritage paintings

Article References: Noori Shaker, B., Raheem Ali, L., Namel, A. T., Ansari, S., Shehada, D., Atalla, S., & Jaafar Hussain, A. (2026). AOViT: an adaptive optimized vision transformer model for the classification of heritage paintings. Frontiers in Artificial Intelligence, 9, Article 1950812. https://doi.org/10.3389/frai.2026.1950812

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1950812

Keywords: vision transformer, artificial intelligence, machine learning, heritage paintings, Persian miniatures, self-attention, AdamW optimizer, Grad-CAM, artistic style classification, digital art history, AOViT, adaptive