Coarse-to-Fine Alignment and Reliability-Driven Fusion for Multispectral Object Detection
Ye Li, Haotian Ning, Tong Xiao, Xiaoke Su, Weiwei Zhang*
School of Computer Science and Artificial Intelligence, Zhengzhou University of Light Industry, Zhengzhou, China
* Corresponding Author: Weiwei Zhang. Email: 
Computers, Materials & Continua https://doi.org/10.32604/cmc.2026.089764
Received 24 July 2026; Accepted 25 August 2026; Published online 20 September 2026
Abstract
To address the challenges of cross-modal spatial misalignment, modality reliability discrepancies, and inadequate supervision in visible-infrared (RGB-IR) multispectral object detection, this paper proposes Coarse-to-Fine Alignment and Reliability-Driven Fusion Network (CFR-Net), a robust framework designed for complex real-world environments. First, we introduce the Scale-aware Offset Alignment and Residual (SOAR) module. By employing a coarse-to-fine offset estimation mechanism, SOAR effectively rectifies global displacements and compensates for local residual misalignments, while leveraging high-frequency residual information to enhance fine-grained representations in small-object regions. Based on the aligned features, the Reliability-aware Quality-guided Fusion (RQF) module is further designed, which adaptively adjusts the fusion weights of RGB and IR features according to region-level modality quality and cross-modal conflict intensity, thereby enhancing reliable complementary information and suppressing low-quality or conflicting responses. Furthermore, an Alignment-aware Fusion Consistency (AFC) Loss is formulated to provide explicit process-level supervision by jointly optimizing feature alignment, fusion consistency, and hard-sample detection, thereby ensuring a more principled learning process that extends beyond standard detection objectives. Experimental results on the DroneVehicle and M3FD datasets demonstrate that a mean Average Precision at an intersection over union (IoU) threshold of 0.50 (mAP50) of 82.1% and 85.7% is achieved by CFR-Net, with a compact architecture of only 6.5M parameters and 16.9 giga floating-point operations (GFLOPs). Consequently, CFR-Net provides a robust and computationally efficient solution for multispectral object detection, achieving competitive performance with minimal architectural overhead.
Keywords
Multispectral object detection; RGB-IR feature fusion; cross-modal alignment; reliability-aware fusion
Open Access
Login
Register
Submit a Paper
Propose a Special lssue
Download PDF
Downloads
Citation Tools
