Image Dehazing using DDPM

Due to the presence of turbid mediums such as water droplets, particles, and sand, haze, smoke, or fog can obscure the scene and result in unclear images. The goal of image dehazing is to improve image clarity, not only to provide better visualization but also to enhance performance in object detection. This project is based on DehazeDDPM. The model utilizes U-Net to extract features for the image formation model and then employs a diffusion model to restore the clear image. Additionally, we incorporate a 2D discrete wavelet module to extract new features and enhance the learning pipeline for improved results. The architecture is shown below:

alt text


We test our model on the haze datasets NH_Haze and Dense_Haze. The results are shown in the table.

NH_Haze Dense_Haze 
PSNRSSIMPSNRSSIM
17.1280.5878715.4070.4254