SRNet: Spectral Rebalancing Network for Robust RGB–Infrared Object Detection Against Modality Anomalies

RGB–infrared object detection relies on complementary information from two modalities, yet it implicitly assumes that the imaging quality of both stays stable. In practice, factors such as component aging and sensor gain drift can degrade one modality, a condition termed a modality anomaly. This work addresses the amplitude-attenuation regime of this condition and evaluates two representative cases: darkening of the visible image and flattening of infrared contrast. Fusion strategies trained on clean data are tied to the healthy statistics of both modalities. Once one side degrades, corrupted features propagate through fixed fusion weights and cause a disproportionate loss in accuracy. To solve this problem, this communication proposes a Spectral Rebalancing Network (SRNet) that restores robustness when either modality is degraded. The network preserves phase and adjusts amplitude in the frequency domain, and it contains two complementary modules: Cross-modal Amplitude Band Swapping (CABS), a training-only augmentation that reduces spectral overfitting at zero inference cost, and Band-wise Spectral Adaptive Fusion (BSAF), which replaces element-wise addition with content-adaptive spectral rebalancing. On M3FD and its degradation benchmarks, SRNet achieves the highest clean-data mAP, and it keeps the best accuracy among the compared methods at every degradation level on both sides. The gap between its accuracy in the two degradation directions is also the smallest.