FDMNet: frequency-domain modulation network for robust object detection in hazy aerial imagery

To achieve object detection in UAV vision, an object detector is crucial, but haze seriously affects detector performance by physically suppressing high-frequency information, which makes it difficult to detect tiny objects. Traditional “dehazing-then-detection” paradigms are restricted to inconsistencies in tasks and restoration effects. The article provides Frequency-Domain Modulation Network (FDMNet), which is an aerial non-explicit-restoration haze-resistant object detector. Following a physical prior, FDMNet constructs an estimate-compensate architecture. A Frequency Domain Modulation (FDM) module is one that expressly estimates the loss of spectral information, and a dynamic LMF-Kernel adaptively restores the loss of mid- and high-frequency discriminative information. To be effective, we present a physical-semantic consistency loss strategy, with frequency-domain consistency loss guaranteeing physical accuracy and prompt distillation loss guaranteeing semantic consistency. We do not have enough datasets, and, therefore, we build Hazy-DOTA, Hazy-DroneVehicle, and a real-life UAV test set. Through extensive experimentation, FDMNet achieved a 7.9% improvement in mAP scores on the HazyDet dataset compared to baseline models, alongside respective gains of 9.9% and 11.5% on the Hazy-DOTA and Hazy-DroneVehicle datasets. Furthermore, it attained state-of-the-art performance relative to several advanced algorithms, balancing both accuracy and robustness.