A single image dehazing algorithm combining adaptive dark channel prior and lightweight CNN
Abstract
Fog and haze degrade outdoor imaging quality and severely impair autonomous driving, surveillance and UAV inspection tasks. To address bright-region color distortion of traditional physical methods, and poor generalization/high complexity of deep learning dehazing approaches, this paper proposes a hybrid single image dehazing algorithm combining Adaptive Dark Channel Prior (ADCP) and lightweight CNN. It adopts a brightness-variance dual-threshold sky segmentation method to significantly reduce DCP's bright-region distortion, designs a depthwise separable convolution-based lightweight CNN with SE attention for transmission map refinement, and uses an improved atmospheric light estimation strategy to enhance color fidelity. Experiments on RESIDE-SOTS-Outdoor, O-HAZE and our self-built Harbin Extreme Dense Fog Dataset show that the proposed algorithm outperforms DCP, AOD-Net and DehazeFormer in PSNR, SSIM and NIQE, achieving superior generalization under extreme dense haze while balancing accuracy, detail preservation and lightweightness.