Fig. 4
From: Real-world defocus deblurring via score-based diffusion models
Defocus deblurring results recovered from different datasets using different methods. Figures (a1)-(a6) show the reconstruction results using the proposed method, the FFDNet method, the DnCNN method, the CycleGAN method, the defocus image and the real image (GT). The same convention is applied to Figures (b1)-(b6), (c1)-(c6) and (d1)-(d6). Figures (e)-(h) correspond to zoomed-in images of the boxed regions in Figures (a6), (b6), (c6), and (d6), respectively.Ours denotes the proposed defocus deblurring method based on the diffusion model of scoring; FFDNet denotes the fast and flexible denoising method based on convolutional neural networks; DnCNN denotes the deep convolutional neural network image denoising method employed; CycleGAN denotes the method of image reconstruction using unsupervised generative modeling. The input is an defocus image; GT is a real image.
