A Haze Removal Method Based on Depth Information Fusion

Abstract

In foggy weather, color distortion and image fading will occur in the captured images. Single image dehazing is a challenging and uncomfortable problem. In recent years, due to the continuous development of deep neural networks, it is gradually used in image dehazing problems as well, which solves the drawbacks of designing haze related features manually. In this paper, we propose a dehazing algorithm based on depth information fusion. We use the KITTI dataset, which contains real haze-free images of different scenes and some additional information. The additional information includes the LiDAR point cloud information and the original depth image, which can get the rough depth information through image fusion. The rough depth information is trained using backpropagation algorithm to complete the missing depth information, which further enriches the scene depth information. The completed accurate scene depth information is used for dehazing in combination with the atmospheric scattering model. Experiments show that our recovered haze-free images retain more image details than some classical research methods and can be detected with more information during target detection.