RASM-Net: A Reliability-Aware Structural and Multiscale Modeling Network for RGB-T Object Detection in Complex Environments

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Visible–thermal (RGB-T) object detection improves robustness in complex environments, such as low-light and densely occluded forest scenes, by integrating complementary information from the two modalities. It has therefore become a key technology for all-weather perception. However, existing methods generally treat the contribution of each modality as fixed or globally consistent, overlooking that modality reliability varies dynamically across scenes and spatial regions. Moreover, fused features may suffer from weakened local structure and cross-scale semantic misalignment during propagation, leading to missed detections and localization errors for small objects and objects with weak boundaries. To address these issues, we propose RASM-Net, a Reliability-Aware Structural and Multiscale Modeling Network that coordinates three complementary stages of the feature-propagation pathway: cross-modal fusion, intra-scale structural modeling, and cross-scale feature aggregation. Dual-Path Reliability-Context Fusion (DRCF) suppresses degraded-modality noise through global and local reliability calibration. Adaptive Kernel-Axial Fusion (AKAF) combines dynamic aggregation with axial positional awareness to preserve local structure. The Multiscale Target-Level Adaptive (MSTA) detection head aligns and aggregates adjacent-scale features before prediction. On the public ODinMJ, M3FD, and FLIR-Aligned datasets, RASM-Net achieves mAP50 values of 96.2%, 81.4%, and 76.4%, respectively, improving upon the direct-concatenation baseline by 2.4%, 3.5%, and 2.0%. The model contains 6.11 M parameters and requires 13.05 GFLOPs. These results show that coordinating reliable fusion, structure preservation, and cross-scale aggregation along the feature-propagation pathway improves detection performance at low model complexity. The proposed approach provides an effective solution for all-weather object perception in nighttime surveillance, autonomous driving, and unmanned aerial vehicles.