Attention-Guided Extraction and Analytical Modeling of Stylistic Features in Oil Paintings

This study proposes an attention-guided framework for extracting and analytically modeling stylistic features in oil paintings, addressing limitations of traditional global descriptor methods that fail to capture local stylistic nuances. Central to the framework is the Neural Palette Encoder (NPE), a multi-branch deep representation model designed to disentangle style from content. Each branch targets distinct stylistic dimensions—such as chromatic distribution, brushstroke directionality, and texture patterns—using embedded attention mechanisms to focus on stylistically significant regions. The NPE incorporates contrastive learning, patch-level coherence constraints, and a Style Perturbation Module to ensure robustness against variations in pigment texture, color shifts, and stroke deformation. Complementing NPE, the Contextual Texture Disentanglement (CTD) module enables structured, context-aware interpretation of painterly styles. CTD models local and global stylistic coherence through a style-aware graph, refining patch embeddings via graph attention networks and top-down modulation. It integrates temporal regularization to account for style evolution within an artist’s oeuvre and supports comparative analysis by balancing global stylistic vectors with localized texture correspondences. The approach ensures interpretability through symbolic archetype alignment and enables fine-grained attribution, influence detection, and stylistic mapping. Experiments on multiple curated oil painting datasets demonstrate significant improvements over state-of-the-art baselines in classification accuracy, robustness, and interpretability. The integration of NPE and CTD not only enhances discrimination between artists and periods but also aligns computational analysis with human perceptual strategies, offering a scalable and transparent solution for art historical research and digital heritage applications.