Article: Learning emotional representations in oil paintings via knowledge distillation and multi-granularity semantic alignment Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.85 pp.70 - 94 Abstract: Oil painting emotion recognition faces two challenges: the excessive number of model parameters makes deployment difficult, and the granularity of emotion semantic understanding is coarse. To address this issue, this paper proposes a lightweight representation learning method that integrates knowledge distillation and multi-granularity semantic alignment. The teacher network guides the lightweight student network to learn, and performs feature alignment at two granularities: global image semantics and local emotional attributes, enabling the compact model to accurately capture the subtle emotions in the paintings while maintaining efficiency. Experiments show that, with 84.7% reduction in parameters, the classification accuracy only drops by 1.2%, and the normalised discounted cumulative gain for emotion intensity ranking reaches 0.893, significantly outperforming the lightweight baseline models of the same scale. This research provides a technical solution that balances efficiency and accuracy for the practical implementation of emotion computing in the field of digital humanities. Inderscience Publishers - linking academia, business and industry through research

Title: Learning emotional representations in oil paintings via knowledge distillation and multi-granularity semantic alignment

Authors: Renjiao Wang

Addresses: Academy of Fine Arts, Nantong Normal College, Nantong, 226000, China

Abstract: Oil painting emotion recognition faces two challenges: the excessive number of model parameters makes deployment difficult, and the granularity of emotion semantic understanding is coarse. To address this issue, this paper proposes a lightweight representation learning method that integrates knowledge distillation and multi-granularity semantic alignment. The teacher network guides the lightweight student network to learn, and performs feature alignment at two granularities: global image semantics and local emotional attributes, enabling the compact model to accurately capture the subtle emotions in the paintings while maintaining efficiency. Experiments show that, with 84.7% reduction in parameters, the classification accuracy only drops by 1.2%, and the normalised discounted cumulative gain for emotion intensity ranking reaches 0.893, significantly outperforming the lightweight baseline models of the same scale. This research provides a technical solution that balances efficiency and accuracy for the practical implementation of emotion computing in the field of digital humanities.

Keywords: knowledge distillation; KD; multi-granularity semantic alignment; oil painting sentiment analysis; sentiment computing.

DOI: 10.1504/IJICT.2026.155483

International Journal of Information and Communication Technology, 2026 Vol.27 No.85, pp.70 - 94

Received: 05 Mar 2026
Accepted: 14 May 2026
Published online: 03 Aug 2026 *

Title: Learning emotional representations in oil paintings via knowledge distillation and multi-granularity semantic alignment

Keep up-to-date