Joint Deraining and Dehazing Using a CNN with Dark Channel Prior and Atmospheric Light Hybrid Model for Robust Image Restoration

Rain, fog and haze severely downgrade visibility conditions, making it difficult for vision-based systems, such as autonomous driver and traffic monitoring systems to function, as they not only compromise the quality of the image, but introduce artifacts.Improving the robustness of these systems in such scenarios is imperative for road safety and dependable decision making.In this work, we propose a new CNN for concurrent image deraining and dehazing, which is built by unifying the power of deep learning with the main concepts from image processing for better traditional image processing.The most remarkable novelty of the method presented in the paper is the fact that it benefits from the estimation of the so-called Dark Channel Prior (DCP), a hand-crafted prior that tells the CNN what a clear image looks like whenever the input is affected by rain streaks as well as haze, since both these atmospheric phenomena occur in outdoor scenarios.The approach thus effectively provides a more robust solution to these complex tasks by combining data-driven learning with handcrafted priors.A novel Atmospheric Light Condition Hybrid model is presented, which exploits the advantages of K-means intensity clustering, CNN-based processing and saliency maps to improve the atmospheric light estimation accuracy, a crucial step in the haze removal process.It also makes corrections for errors in atmospheric light estimation, giving the final restoration a better quality.The architecture of the proposed CNN is a multi-stage decomposition, with each stage tailored to a specific degradation item: rain streak separation, haze-free region recovery, and clean image reconstruction.Large amounts of experimental data show that PSNR and SSIM of 28.14 dB and 0.901 are achieved under light rain in the Rain L dataset, and PSNR of 25.99 dB and SSIM of 0.873 are obtained under moderate haze by the I-Haze dataset, outperforming significantly compared with existing state-of-the-art methods.Besides, our model keeps robustness in more heavy degradation: PSNR=22.097dB,SSIM=0.813onRain 1400 dataset and PSNR=21.90,SSIM=0.812 on NH-Haze datasets.These results demonstrate the model's robustness in processing diverse rain and haze conditions, leading to considerable improvements in restoring images over the best previously published results.The approach presents a new state of the art for image restoration tasks in harsh weather conditions.