Inter-Scale Sure-Let Image Restoration with Deep Unrolled Image Prior

This study extends a self-supervised image denoising technique proposed by the authors to a more general image restoration method. The previous work was inspired by Ulyanov’s deep image prior (DIP) method. DIP uses a deep convolutional network as an image prior to generate a restored image from a random one, which brings an advantage of no training data requirement. However, one problem arises that the interpretability is low and it is not trivial to explain the need for the random input. In the previous work, it is shown that Stein’s unbiased risk estimator (SURE) with Monte-Carlo computation can be used to train a denoising network. This approach allows us to interpret the random input requirement. The framework trains an image denoiser instead of an image generator. This work extends the discussion to more general image restoration problem by introducing a deep unrolling approach to reflect the measurement process. The new framework uses the Luisier’s inter-scale linear expansion of the thresholding (LET) for exploiting the cross-correlation among extracted features. Some simulation results of image restoration show the significance of the proposed method.