Ulyanov D, Vedaldi A, Lempitsky V. Deep image prior[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 9446-9454.
1. Overview
目前深度学习都是learning-based方式,论文将深度学习作为learning-free方式
- used as handcrafted priors to solve inverse problems, such as denoising, super-resolution, inpainting, restore image based on flash-no flash image
- generator network is sufficient to capture low-level image statistics proir
- bridge the gap between learning-based and laerning-free
- first study to investigates the prior captured by deep convolutional generative networks independently of learning the network parameters from images
1.1. Loss Function
- input. random noise, usually initialize randomly and keep it fixed
- gt. noise image
1.2. Converge
- 图像的收敛比噪声快
1.3. 应用
Denoising
Super-Resolution
- input. random noise
- gt. H x W
- output. tH x tW, and then downsampled H x W
- Inpainting
- output. x
- gt. x0, image with missing pixel
- m. binary mask
skip-connection对该类任务的效果较差
- Natural Pre-image
pre-image
loss function
- Flash-no Flash Reconstruction