Image deblurring method using denoising diffusion probabilistic model

Abstract

Since the blurring artifact can degrade the quality of the images as well as the performance of other important image-related tasks, studies on deblurring are actively underway. Accordingly, existing image deblurring methods have achieved high performance based on CNN-based models and GAN-based models. However, they have problems such as over-smoothing because of minimization of pixel loss and un-convergence in training. In this research, we propose an image deblurring method using a Denoising Diffusion Probabilistic Model. This model destroys the distribution of images by iteratively adding noise to clean images in the diffusion process, thereby finally becoming Gaussian distribution. Then, the noise is estimated and the estimated noise is subtracted from the input Gaussian distributed noise to restore the distribution of clean image. Based on this model, we generate deblurred images from blurry images which are the conditional input of the noise estimation model. The proposed model has particular strengths in terms of perception as it learns image distribution. It is demonstrated that our proposed method achieves high performance in perceptual evaluations than existing image-deblurring methods.