From learning models of natural image patches to whole image restoration

By Daniel Zoran, Yair Weiss. 2011

1,572 citations recorded by the source corpus.

Subfields: Computational Mechanics, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Mechanics

In Computer Vision and Pattern Recognition

Citations in the corpus

Neural Autoregressive Distribution Estimation (2016, methodology): “First, we learned the means of the MoG components, while Zoran and Weiss (2011) ﬁxed them to zero.”
Generative Visual Manipulation on the Natural Image Manifold (2016, methodology): “A common strategy is to learn local filter or patch models, such as Principal Components, Independent Components, Mixture of Gaussians or wavelet bases [14,15,16].”
Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks (2015, background): “Very large Gaussian mixture models [34] and sparse coding models of image patches [31] can also be used but suffer the same problem.”
Generative Image Modeling Using Spatial LSTMs (2015, methodology): “Generative image models are also useful in more traditional applications such as image reconstruction [33, 35, 49] or compression [47].”
RNADE: The real-valued neural autoregressive density-estimator (2013, methodology): “The best RNADE test log-likelihood is, on average, 0.7 nats per patch lower than Zoran and Weiss’s MoG, which had a different training procedure than our mixture of Gaussians.”
Masked Autoregressive Flow for Density Estimation (2017, background): “For instance, density estimators can learn suitable priors for data from large unlabelled datasets, for use in standard Bayesian inference [39].”
Learning Generative Models with Visual Attention (2013, methodology): “Therefore, generative models have found most success in learning to model small patches of natural images and objects: Zoran and Weiss [4] learned a mixture of Gaussians model over 8×8 image patches; Salakhutdinov and Hinton [2] used 64×64 centered and uncluttered stereo images of toy objects on a c”
Examining the Impact of Blur on Recognition by Convolutional Networks (2016, methodology): “We report performance of the original network (applied at scale 256) on sharp images, those blurred with D8, and then deblurred using [28] with knowledge of the blur kernel.”
Pixel Recursive Super Resolution (2017, methodology): “One approach to this problem is to define the conditional distribution of the output pixels jointly as a multivariate Gaussian mixture [52] or an undirected graphical model [10].”
Deep Mixtures of Factor Analysers (2012, methodology): “As confirmed by Zoran & Weiss (2011), a better prior almost certainly leads to a better signal to noise ratio of the restored image.”
Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising (2016, methodology): “It can be seen that the effective patch size used in EPLL [33] is the smallest, i.e., 36 × 36 .”
Deep Generalized Unfolding Networks for Image Restoration (2022, methodology): “For instance, [100] integrated Gaussian Mixture prior to HQS.”
Gaussian Conditional Random Field Network for Semantic Segmentation (2016, methodology): “Gaussian conditional random fields: GCRFs [50] are popular models for structured inference tasks like denoising [22, 43, 44, 50, 56], deblurring [43, 44, 56], depth estimation [29, 42], etc.”
Maintaining Natural Image Statistics with the Contextual Loss (2018, methodology): “on \Facts are stubborn things, but statistics are pliable." | Mark Twain Maintaining natural image statistics has been known for years as a key factor in the generation of natural looking images [1,2,3,4,5]. With the rise of CNNs, the utilization of explicit image priors was replaced by Generat”
Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring (2016, methodology): “Final latent image estimation is performed with conventional optimization method [31].”
Proximal Deep Structured Models (2016, result): “We compare against a number of recent state-of-the-art techniques [7, 46, 22, 33, 32].”
$L_0$ -Regularized Intensity and Gradient Prior for Deblurring Text Images and Beyond (2017, methodology): “For fair comparisons, all the final deblurred results are generated by the same non-blind deconvolution method [50], and the error ratio metric [13] is used for evaluation.”
Solving inverse problems using data-driven models (2019, methodology): “Examples include K-SVD, which is a sparse coding approach on image patches (Elad and Aharon 2006), BM3D, which combines sparsity and self-similarity (Dabov, Foi, Katkovnik and Egiazarian 2007), and EPLL, which is a Gaussian mixture model of image patches (Zoran and Weiss 2011).”