Non-blind Deblurring Using Probabilistic Models and Spatial Adaptive Restoration

The core concept of image deblurring is to establish a prior model and formulate the deblurring process as an optimization problem by using the statistical properties of gradients and noise. In usual, a user-defined parameter is introduced to control the trade-off between the sharpness of the deblurring result and the robustness to noise. However, using a single variable is difficult to make the deblurring algorithm adapt to different types of images. In this work, we aim to overcome this limitation. The criteria to determine different regions can be obtained through noise estimating and construct a space-variant noise model. Utilizing different statistical characteristics of noise, we can approximate their distribution using various hyperLaplacian models. Thus, we establish two lookup mechanisms. The final result interpolates the result of different regions to obtain a smoother outcome. Additionally, for the significantly degraded image with a low sign-to-noise ratio (SNR), we apply an adaptive denoiser in the frequency domain to stabilize the deblurred image.