DmitryUlyanov/deep-image-prior
Deep Image Prior
This project is an unsupervised image restoration tool that uses a convolutional neural network as a structural prior to reconstruct images from noisy or incomplete data. It functions as a neural network image prior, utilizing the inherent biases of the network architecture to restore pixels without the need for a pre-trained dataset or external learning.
The system performs zero-shot image restoration by treating the network architecture itself as a regularization term. It uses a randomly initialized encoder-decoder structure and iterative gradient descent to minimize pixel-wise loss, recovering the natural image structure through self-supervised reconstruction.
The tool covers several image recovery domains, including image denoising, super-resolution, and general image restoration. It also provides capabilities for blind image deconvolution to recover sharp images from blurred versions.
Features
- Image Restorers - Provides a convolutional neural network based tool for unsupervised image denoising and super-resolution.
- Structural Priors - Implements a convolutional neural network as a structural prior to restore images without needing external training datasets.
- Self-Supervised Reconstruction - Performs image restoration by treating the neural network architecture as a self-supervised regularization term.
- Image Restoration Models - Repairs corrupted image data and fills missing pixels by optimizing a reconstruction network.
- Unsupervised Restoration Models - Recovers corrupted pixels using an unsupervised approach that requires no pre-trained dataset.
- Image Super Resolution Models - Increases image resolution and restores fine details by leveraging the structural properties of a convolutional network.
- Iterative Image Optimizers - Refines reconstructed image pixels through iterative gradient descent to minimize pixel-wise loss.
- Generative Priors - Leverages the inherent bias of convolutional layers as a structural prior for image reconstruction.
- Zero-Shot Restoration - Performs zero-shot image restoration by applying a neural network architecture as a prior without pre-trained datasets.
- Image Denoising - Removes unwanted noise from digital images using a neural network prior without training data.
- L1 Pixel Loss - Minimizes the L1 pixel-wise difference between the synthesized output and the corrupted target image.
- Encoder-Decoder Architectures - Utilizes an encoder-decoder structure to capture multiscale spatial information for image reconstruction.
- Weight Initialization - Uses random weight initialization to ensure the network recovers natural image structures through the optimization process.
- Blind Image Deconvolution - Provides capabilities for blind image deconvolution to recover sharp images from blurred versions.
- Joint Denoising and Super-Resolution - Combines image denoising and super-resolution capabilities using a randomly initialized network prior.
- Computer Vision Models - Restores images using neural networks without requiring prior learning.
- Computer Vision Research - Unsupervised image restoration using neural network architecture priors.
- Generative Models - Deep learning-based image restoration without pre-training.
- Image Inpainting Models - Unsupervised image restoration using deep generative priors.
- Image Transformation - Deep image prior for restoration tasks.
- Model Implementations - Image restoration without learning.
- Non-Blind Deblurring - Uses deep image priors for various restoration tasks without pre-training.
