Fog Traffic Image Simulation Optimization Algorithm Based on Unet++

This study aims to improve the challenge of achieving accurate target detection accuracy under fog conditions. Aiming at the scarcity of foggy data sets, we propose a fog traffic image simulation optimization algorithm based on UNet++. The algorithm consists of several steps: the relative depth map is generated by the self-supervised monomeric depth estimation network, and then the absolute depth map is initially derived by applying dense geometric constraints for scale recovery. Secondly, the Unet ++ image segmentation network is used to segment the sky region of the original image and optimize the initial absolute depth map. Then, the visibility of the simulated image is defined and the transmittance map is generated. Then dark channel maps are used to distinguish sky regions and estimate atmospheric light values. Finally, atmospheric scattering model is used to generate fog simulation images under specific visibility conditions. The experimental results show that compared with before optimization, the optimized fog simulation algorithm without reference image quality evaluation score is significantly improved. This suggests that its non-structural similarity (NSS) features are closer to those of natural fog. The fog simulation method proposed in this paper can realize clear image conversion under natural environment, and solve the problem of lack of fog image data set and incomplete visibility data.