.png){kind=icon}
.png){kind=icon}
Return
Progressive knowledge evolution under multi-granularity knowledge distillation for infrared action recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2026.115128.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, infrared action recognition, an important application of artificial intelligence, has gained increasing attention due to its robustness under low-light and privacy-sensitive conditions. Due to the great achievement of visible light-based action recognition, most existing infrared action recognition methods are adapted from visible light-based approaches and primarily rely on visual features. However, infrared data lacks detailed information such as texture and color found in visible light data, which leads to suboptimal performance of these methods. Moreover, current multi-modal recognition methods typically require multi-modal inputs during inference, limiting their practicality in real-world infrared-only scenarios. Thus, we propose a homologous Knowledge Evolving Teacher-Groups for Infrared Action Recognition via Multi-Granularity Knowledge Distillation. The framework constructs a teacher group from a visual knowledge source and progressively evolves knowledge along visual representations, pose structures, and semantic descriptions, thereby providing cross-modal guidance to an infrared-only student network to learn comprehensive and highly discriminative action representations, resulting in high-quality infrared action recognition. The multi-granularity distillation strategy, which ensures sample-, class-, and decision-level, strengthens the overall representation capability of the infrared modality in feature modeling, class discrimination, and boundary awareness by avoiding limited knowledge representation and overfitting. As a result, our method achieves a student network with great generalization and robustness for infrared action recognition in complex scenarios. Our method outperforms existing state-of-the-art methods by achieving 89.37% and 90.19% accuracy on two infrared datasets.
Keywords:
infrared action recognition
knowledge distillation
multi-granularity
teacher-student framework
action recognition
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W
Organization
No organization information available