PE-Det : Prior-Guided visible preconditioning and routed expert fusion for robust infrared-visible object detection

Liu, Y. and Cai, S. and Guo, X. and Geng, Q. and Ling, X. (2026) PE-Det : Prior-Guided visible preconditioning and routed expert fusion for robust infrared-visible object detection. Expert Systems with Applications, 321: 132275. ISSN 0957-4174

Text (PE-Det_Prior-Guided_visible_preconditioning_and_routed_expert_fusion_for_robust_infrared-visible_object_detection)
PE-Det_Prior-Guided_visible_preconditioning_and_routed_expert_fusion_for_robust_infrared-visible_object_detection.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (2MB)

Abstract

Infrared–visible object detection in adverse conditions remains challenging because severe degradation in the visible modality, including low illumination, haze scattering, and contrast collapse, causes cross-modal inconsistency and undermines the reliability of fixed fusion strategies across scenes and object scales. To address this issue, we propose PE-Det, a degradation-aware adaptive fusion framework that improves detection robustness through prior-guided visible preconditioning, routed expert fusion, and multi-scale aggregation. Specifically, the PVP module preconditions the visible input by stabilizing photometric statistics and enhancing structure-relevant cues, thereby providing more reliable features for subsequent cross-modal interaction. An expert fusion pool with routing-based selection is then introduced across pyramid levels, allowing the model to adapt its fusion behavior to scene complexity and scale-dependent cross-modal discrepancies. In addition, the GS-SSFF neck strengthens cross-scale interactions and produces detection-oriented multi-scale representations. To further improve localization under degradation, conventional IoU-based box regression is replaced with the proposed CFI-MPD-IoU, which enhances the stability and consistency of bounding-box optimization. Extensive experiments on FLIR and M3FD show that PE-Det consistently outperforms representative single- and multi-modal YOLO baselines and a broad range of publicly available fusion-driven detectors. Cross-dataset evaluations further verify its robustness to unseen degradations and domain shifts. The source code is publicly available at https://github.com/601140736/PE-Det.