Enhancing Thermal Image Object Detection using Spatial Edge-aware Attention and Self-supervision Pretext

Gaeul Han, Thangarajah Akilan
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:538-551, 2026.

Abstract

Thermal cameras offer robust sensing for object detection in low-visibility driving conditions, but thermal images often suffer lower resolution and weaker object boundaries than RGB imagery. This paper presents SEA-YOLO-E (Spatial Edge Attention YOLO-E), an enhanced single-modality thermal object detector that integrates a SEA mechanism and semi-supervised learning to overcome these challenges. First, we introduce the SEA-YOLO architecture, which embeds an Edge Extractor and a novel SEA module into a YOLOv8 backbone to emphasize object boundaries and improve detection accuracy in thermal domains. Bases on it, we extend SEA-YOLO with a semi-supervised learning paradigm: a self-supervised rotation prediction pretext task leverages unlabeled infrared images to learn general feature representations, and synthetic thermal data mitigates class imbalance in training. The proposed two-phase training (self-supervised pretraining followed by supervised fine-tuning) significantly boosts detection performance. Experiments on multiple thermal driving datasets demonstrate that SEA-YOLO-E achieves state-of-the-art results, with improvements of up to 9–12% in mAP over existing detectors. Notably, our edge-enhanced attention and rotation-pretrained model outperforms recent multi-modal RGB-thermal detectors while using only thermal input.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-han26a, title = {Enhancing Thermal Image Object Detection using Spatial Edge-aware Attention and Self-supervision Pretext}, author = {Han, Gaeul and Akilan, Thangarajah}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {538--551}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/han26a/han26a.pdf}, url = {https://proceedings.mlr.press/v318/han26a.html}, abstract = {Thermal cameras offer robust sensing for object detection in low-visibility driving conditions, but thermal images often suffer lower resolution and weaker object boundaries than RGB imagery. This paper presents SEA-YOLO-E (Spatial Edge Attention YOLO-E), an enhanced single-modality thermal object detector that integrates a SEA mechanism and semi-supervised learning to overcome these challenges. First, we introduce the SEA-YOLO architecture, which embeds an Edge Extractor and a novel SEA module into a YOLOv8 backbone to emphasize object boundaries and improve detection accuracy in thermal domains. Bases on it, we extend SEA-YOLO with a semi-supervised learning paradigm: a self-supervised rotation prediction pretext task leverages unlabeled infrared images to learn general feature representations, and synthetic thermal data mitigates class imbalance in training. The proposed two-phase training (self-supervised pretraining followed by supervised fine-tuning) significantly boosts detection performance. Experiments on multiple thermal driving datasets demonstrate that SEA-YOLO-E achieves state-of-the-art results, with improvements of up to 9–12% in mAP over existing detectors. Notably, our edge-enhanced attention and rotation-pretrained model outperforms recent multi-modal RGB-thermal detectors while using only thermal input.} }
Endnote
%0 Conference Paper %T Enhancing Thermal Image Object Detection using Spatial Edge-aware Attention and Self-supervision Pretext %A Gaeul Han %A Thangarajah Akilan %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-han26a %I PMLR %P 538--551 %U https://proceedings.mlr.press/v318/han26a.html %V 318 %X Thermal cameras offer robust sensing for object detection in low-visibility driving conditions, but thermal images often suffer lower resolution and weaker object boundaries than RGB imagery. This paper presents SEA-YOLO-E (Spatial Edge Attention YOLO-E), an enhanced single-modality thermal object detector that integrates a SEA mechanism and semi-supervised learning to overcome these challenges. First, we introduce the SEA-YOLO architecture, which embeds an Edge Extractor and a novel SEA module into a YOLOv8 backbone to emphasize object boundaries and improve detection accuracy in thermal domains. Bases on it, we extend SEA-YOLO with a semi-supervised learning paradigm: a self-supervised rotation prediction pretext task leverages unlabeled infrared images to learn general feature representations, and synthetic thermal data mitigates class imbalance in training. The proposed two-phase training (self-supervised pretraining followed by supervised fine-tuning) significantly boosts detection performance. Experiments on multiple thermal driving datasets demonstrate that SEA-YOLO-E achieves state-of-the-art results, with improvements of up to 9–12% in mAP over existing detectors. Notably, our edge-enhanced attention and rotation-pretrained model outperforms recent multi-modal RGB-thermal detectors while using only thermal input.
APA
Han, G. & Akilan, T.. (2026). Enhancing Thermal Image Object Detection using Spatial Edge-aware Attention and Self-supervision Pretext. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:538-551 Available from https://proceedings.mlr.press/v318/han26a.html.

Related Material