Semi-supervised single-image dehazing based on spatial-channel feature enhancement

Visibility reduction in haze-laden environments significantly hinders object discernment, presenting a substantial challenge in image processing. Current supervised dehazing methods are heavily reliant on the quality and diversity of their training datasets, which limits their generalization capabilities and incurs high costs due to the need for accurately paired training images. To address these limitations, this paper introduces a semi-supervised single-image dehazing network that leverages both synthetic and real-world hazy images during training. This mixed training approach enhances the model’s applicability to diverse real-world conditions and improves its robustness against varying haze densities. Our methodology incorporates a novel spatial-channel feature enhancement module that optimally processes images with uneven feature distributions and significant interference, thus maintaining integrity in feature representation. We evaluate our approach on multiple public benchmark datasets, the proposed method achieved a PSNR score of 20.93, SSIM scores of 0.724 on the NH-HAZE dataset, where it outperforms existing supervised and unsupervised methods in terms of generalization to both synthetic and authentic hazy images.