A style aware attention transformer for fine grained interior design style recognition

Fine-grained interior design style recognition is challenging due to the subtle, context-dependent nature of stylistic cues, which are often entangled with dominant semantic objects. While convolutional neural networks effectively capture local textures, they lack global reasoning, whereas Vision Transformers model spatial relationships but do not explicitly encode perceptual style attributes. To address these limitations, we propose SAT $$^{3}$$ (Style-Aware Attention Transformer) that integrates four synergistic components: (i) a Style Attention Embedding that combines a CNN (Convolutional Neural Network) stem with a Transformer encoder enhanced by a Style Attention Module (SAM) to emphasize texture and color-related cues; (ii) an auxiliary self-supervised learning (SSL) branch leveraging masked autoencoding and contrastive learning to enrich style-invariant representations; (iii) a joint optimization strategy that balances supervised and self-supervised objectives for improved robustness; and (iv) an efficient inference pathway that retains a style-sensitive encoder while removing auxiliary components. Extensive experiments on interior design datasets demonstrate that SAT $$^{3}$$ achieves the best performance among the baselines evaluated in this study. Additional analyses indicate that the learned representations are more aligned with stylistic attributes and less dependent on semantic content, leading to improved generalization, particularly under limited annotation settings. These results highlight the effectiveness of the proposed approach for fine-grained style recognition and suggest its broader applicability to other style-driven visual understanding tasks.