Single image dehazing via dual-branch spatial network with cross-attention knowledge distillation

Single image dehazing is a typical ill-posed image restoration problem, and existing methods struggle to balance global feature extraction and computational efficiency. This paper proposes a dual-branch spatial network fusing frequency domain analysis and cross-attention knowledge distillation. It decouples features into amplitude-phase dual branches via Fast Fourier Convolution to realize low-frequency energy calibration and phase structural fidelity respectively. A Dual-Driven Prior Gating Network based on knowledge distillation and Kolmogorov-Arnold Network (KAN) is designed to distill channel-wise prior knowledge through cross-attention and generate adaptive gating weights via KAN. A spatial-spectral joint multi-loss function is constructed for network training. Experiments on RESIDE and IO-HAZE show that the method outperforms mainstream algorithms significantly in quantitative metrics and visual effects, verifying its effectiveness and superiority.