Deep Image Prior

This paper introduces a new approach to image restoration that doesn't require learning from large datasets. The core idea is that the structure of convolutional neural networks (CNNs) itself acts as a prior, favoring natural-looking images during the restoration process. By fitting a CNN to a corrupted image, the network gradually recovers the clean image, effectively acting as a hand-designed prior that can be used for various tasks like denoising, super-resolution, and inpainting without needing pre-training.