What makes a good model of natural images?

By Yair Weiss, William T. Freeman. 2007 IEEE Conference on Computer Vision and Pattern Recognition, 2007

364 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Computational Mechanics

View Paper: https://www.semanticscholar.org/paper/bee150f94d61d09e08d858605af36d1b96d2bc52

How researchers see this paper

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

In Artificial Intelligence

In Computational Mechanics

Citations in the corpus

Learning Multiscale Sparse Representations for Image and Video Restoration (2008, background): “Other results on learning single-scale image dictionaries include, for example, [37, 38, 43].”
Fine-Grained Similarity Measurement between Educational Videos and Exercises (2020, background): “Theis et al. [31] used the conditional distribution of a mixture of GSMs (gaussian scale mixtures) [34, 37] to model the distribution of a pixel given its causal neighborhood.”
Internal statistics of a single natural image (2011, background): “, [16, 13, 15]) impose parametric distribution on natural image responses to local filters.”
CG2Real: Improving the Realism of Computer Generated Images Using a Large Collection of Photographs (2011, background): “Others have focused solely on learning a model for the statistics of natural images [Weiss and Freeman 2007; Roth and Black 2005].”
Image denoising: Can plain neural networks compete with BM3D? (2012, methodology): “Other methods also learn a global image prior on a noise-free dataset, for instance [20, 27, 9].”
Learning optimized MAP estimates in continuously-valued MRF models (2009, methodology): “Recent years have seen the introduction of several new approaches for learning the parameters of continuousvalued Markov Random Field models [4, 11, 15, 20, 9, 12].”
The student-t mixture as a natural image patch prior with application to image compression (2014, background): “The Fields of Experts (FoE) framework is another approach for learning priors that can be used for image processing applications (Roth and Black, 2005; Weiss and Freeman, 2007).”
Real Image Denoising With Feature Attention (2019, methodology): “Model-based algorithms include non-local self-similarity (NSS) [18, 13, 20], sparsity [30, 48], gradient methods [46, 56, 54], Markov random field models [52], and external denoising priors [9, 61, 42].”
Practical Bounds on Image Denoising: From Estimation to Information (2011, background): “The entropy could be calculated if the prior pdf could be ascertained or modeled accurately at all . Although many have proposed various models for natural images [32]–[ 34 ], they are not directly applicable to our case since we consider the patch vectors to be geometrically similar within each clu”
Nonlinear image representation using divisive normalization (2008, background): “Previously, Student’s t has been proposed as an image model mainly for its convenience in computation [23, 18, 29].”
From learning models of natural image patches to whole image restoration (2011, background): “[9, 1, 14]) but they typically constrain the means and covariance matrices during learning.”
Discriminative learned dictionaries for local image analysis (2008, background): “Recent publications have shown that learning non-parametric dictionaries for image representation instead of using off-theshelf ones, can significantly improve image restoration, e.g., [6] [30] [31] [38].”
Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising (2016, background): “Over the past few decades, various models have been exploited for modeling image priors, including nonlocal self-similarity (NSS) models [1], [2], [3], [4], sparse models [4], [5], [6], gradient models [7], [8], [9] and Markov random ﬁeld (MRF) models [10], [11], [12].”
Blind deconvolution using a normalized sparsity measure (2011, background): “Weiss and Freeman [23] propose a simpler learning scheme for the Fields of Experts model that allows the efficient training of large filters.”
Regularized estimation of image statistics by Score Matching (2010, methodology): “(i) Filters of model with 5× 5× 24 weights learned with CD-1 [21], (j) filters of our model with 5 × 5 × 24 weights, (k) random selection of filters from the Basis Rotation [23] model with 15× 15× 25 weights, (l) random selection of filters from our model with 8× 8× 64 weights.”
Statistics of real-world hyperspectral images (2011, positive): “Our work is motivated by successes in analyzing and modeling the statistical properties of grayscale images [1, 30, 39].”
Who killed the directed model? (2008, methodology): “These have been successfully learnt as priors for natural images [13, 22, 23], stereo depth [16, 15], single image depth [14], object segmentation [2], optical flow [12], intrinsic images [20], etc.”