GenDist: Generative-Driven Knowledge Distillation for Small Object Detection in Remote Sensing Images
IEEE Transactions on Geoscience and Remote Sensing · 2026 · 64 · 1-16
DOI: 10.1109/tgrs.2026.3676015 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Geoscience and Remote Sensing, 2026. Verfasst von Zhuangzhuang Zhou und 5 weiteren Autor:innen. Die Arbeit wurde bisher 4-mal zitiert (Crossref) und nennt 86 Literaturangaben. Gefördert von: Science and Technology Innovation Leading Talent Support Program of Henan Province, Natural Science Foundation of Henan, Science and Technology Research Project of Henan Province.
Abstract (Originalsprache)
Optimal efficiency is always the primary pursuit of the remote sensing community. However, the blurred appearance in small objects causes increasingly deeper and wider detectors, posing a significant challenge for deployment. knowledge distillation (KD), which can shorten the time from image acquisition to interpretation, has become a hotspot. In this article, contrary to traditional feature imitation methods, we propose generative-driven KD for small object detection (GenDist), which reconstructs teacher-like representations from the partially masked student features to achieve knowledge transfer. Masking during training strengthens the student’s ability to generalize in complex scenes. Specifically, GenDist contains three parts: online self-correct mask (O-SCM), dynamic activation global context (DAGC), and sampling-aligned generation (SAG). First, O-SCM is utilized to identify informative masking regions that guide the student to reconstruct the teacher-like features. By integrating an online self-correction mechanism, it adaptively directs generative supervision toward regions that are ignored or misinterpreted by students. Then, DAGC refines teacher features via global context modeling and critical feature selection, facilitating more accurate object detail capture by the student model. Finally, SAG improves the quality of feature reconstruction by introducing the learnable offset sampling, enabling students to learn irregular areas such as the small instances’ edges and texture transitions. Our proposed method finds the sweet spot between precision and speed in an elegant way. GenDist yields stable performance gains for student models across the DIOR, AI-TOD, SODA-D, and SSDD datasets. For the lightweight detectors, it cuts the model’s parameter count by roughly 38.2%, while keeping accuracy loss under 3.9%.
Förderung
- Science and Technology Innovation Leading Talent Support Program of Henan Province
- Natural Science Foundation of Henan
- Science and Technology Research Project of Henan Province
- Key Research Project of Henan Province Higher Education Institutions
- Shenzhen Science and Technology Program
Zitieren
APA
Zhou, Z., Zhao, X., Zhao, X., Zhu, Y., Zhao, Z., & Li, L. (2026). GenDist: Generative-Driven Knowledge Distillation for Small Object Detection in Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 64, 1-16. https://doi.org/10.1109/tgrs.2026.3676015
Vancouver
Zhou Z, Zhao X, Zhao X, Zhu Y, Zhao Z, Li L. GenDist: Generative-Driven Knowledge Distillation for Small Object Detection in Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing. 2026;64:1-16. doi:10.1109/tgrs.2026.3676015
BibTeX
@article{Zhou2026,
author = {Zhou, Zhuangzhuang and Zhao, Xuewu and Zhao, Xuezhuan and Zhu, Yingying and Zhao, Zunjin and Li, Lingling},
title = {GenDist: Generative-Driven Knowledge Distillation for Small Object Detection in Remote Sensing Images},
journal = {IEEE Transactions on Geoscience and Remote Sensing},
year = {2026},
volume = {64},
pages = {1--16},
doi = {10.1109/tgrs.2026.3676015}
}
EUVIMED