Single image defogging via Physical Model and Hyperresolution Reconstruction
Chenghao Zhu, Wei Hong Liu, Pengcheng Yang · 2023
Image defogging is an important computer vision method. The current end-to-end defogging method, Convolutional Neural Networks (CNNs), has achieved significant success, and estimating depth information has become a mainstream method. However, depth information only focuses on the type of object, resulting in texture information loss and edge blur issues. To solve this problem.This article introduces a new defogging network (PHD) based on physical and image reconstruction. We first use a depth estimation network to obtain depth information and learn to physically decompose haze images into two components that conform to the scattering model: transmitted images and atmospheric light. In addition, we combine depth information with prior details of the fog map, learn the areas we need to focus on through the designed attention network, and generate weight matrix parameters to guide defogging. In the fine-tuning stage, we use the image hyper variational algorithm to repair its texture details. The experimental results indicate that compared with existing algorithms, this method has good performance.