Image-to-image machine translation enables computational defogging in real-world images

This paper addresses the challenge of computational defogging using image-to-image (I2I) machine learning models trained on real-world data. We introduce Stereofog, the largest and most diverse dataset to date, comprising 10, 067 paired clear-foggy images captured with a custom-built binocular camera setup. By training a pix2pix I2I model on this dataset, we achieve a Complex Wavelet Structural Similarity Index (CW-SSIM) of 0.76, Multi-scale Structural Similarity Index (MS-SSIM) of 0.7, and Pearson correlation coefficient of 0.4 for defogged images, demonstrating significant improvements in defogging efficacy compared to models trained on synthetic data. The model maintains high performance with a CW-SSIM of 0.95 for low fog density and 0.8 for real data, though it drops to 0.5 at very high fog densities. These results underscore the model’s ability to produce plausible reconstructions under varying fog conditions. This study advances the field by providing a robust, open-source dataset, and demonstrating the practical applicability of open-sourced I2I machine learning models for real-world computational defogging.