Emotion representation and recognition in oil paintings via meta-learning and semantic augmentation

Automated emotion analysis in visual art remains a significant challenge, primarily due to the paucity of annotated data and the profound stylistic and semantic gap between generic image understanding and domain-specific artistic interpretation.This study introduces a novel meta-learning framework enhanced with structured semantic knowledge for few-shot emotion recognition in oil paintings.The proposed model integrates a dual-path architecture: a meta-learning pathway for rapid visual adaptation and a semantic pathway that incorporates contextual art historical knowledge.These pathways are fused through a hierarchical cross-modal attention module, which dynamically aligns visual features with relevant semantic concepts during the learning process.Extensive evaluations on the ArtEmis dataset demonstrate the framework's superior performance, achieving state-of-the-art macro-accuracy of 68.7% (1-shot) and 81.3% (5-shot).The results confirm the model's efficacy in achieving robust, generalisable, and interpretable emotion analysis with limited data, advancing the field of computational art understanding.