The goal is to improve single-image deraining methods for UAV-assisted IoT systems by modeling long-range dependencies and utilizing frequency-domain features.
Proposed SFQMamba framework combining a CNN branch with a Fused Enhance Block and Mamba branch with Spatial-Aware Selective Fusion Block.
Extended 1D State Space Models to the 2D domain for enhanced feature fusion using a scanning mechanism.
Evaluated using Rain13K, Raindrop datasets, and RainVisDrone benchmark.
SFQMamba improved PSNR by 0.12 dB and SSIM by 0.11% compared to TransMamba.
On the RainVisDrone benchmark, improved detection with YOLOv8s: 0.0737 AP, 0.1060 AP50, and 0.0897 AP75.
Effectively removed dense rain streaks while preserving structural and textural details.