On learning optimized reaction diffusion processes for effective image restoration

By Yunjin Chen, Wei Yu, Thomas Pock. Computer Vision and Pattern Recognition, 2015

335 citations recorded by the source corpus.

For several decades, image restoration remains an active research topic in low-level computer vision and hence new approaches are constantly emerging. However, many recently proposed algorithms achieve state-of-the-art performance only at the expense of very high computation time, which clearly limits their practical relevance. In this work, we propose a simple but effective approach with both high computational efficiency and high restoration quality. We extend conventional nonlinear reaction diffusion models by several parametrized linear filters as well as several parametrized influence functions. We propose to train the parameters of the filters and the influence functions through a loss based approach. Experiments show that our trained nonlinear reaction diffusion models largely benefit from the training of the parameters and finally lead to the best reported performance on common test datasets for image restoration. Due to their structural simplicity, our trained models are highly efficient and are also well-suited for parallel computation on GPUs.

Subfields: Computer Vision and Pattern Recognition, Radiology, Nuclear Medicine and Imaging

View Paper: https://www.semanticscholar.org/paper/177bded5ff0c9a621f29ac1e7d920e8a3ee7bab8

How researchers see this paper

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

In Computer Vision and Pattern Recognition

In Radiology, Nuclear Medicine and Imaging

Citations in the corpus

Framing U-Net via Deep Convolutional Framelets: Application to Sparse-View CT (2017, methodology): “…streaking artifacts originated from sparse projection views show globalized artifacts that are difﬁcult to remove using conventional denoising CNNs [13], [14], Authors are with the Department of Bio and Brain Engineering, Korea Advanced Institute of Science and Technology (KAIST), Daejeon 34141,…”
Deep Convolutional Framelets: A General Deep Learning Framework for Inverse Problems (2017, background): “Interestingly, this non-local basis is quite often used in CNNs that do not have a pooling layer.”
Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising (2016, background): “[15], [16] proposed a trainable nonlinear reaction diffusion (TNRD) model which learns a modiﬁed ﬁelds of experts [12] image prior by unfolding a ﬁxed number of gradient descent inference steps.”
Solving inverse problems using data-driven models (2019, methodology): “Other examples include learning coefficients of a PDE via optimal control (Liu, Lin, Zhang and Su 2010), as well as deriving CNN architectures motivated by diffusion processes (Chen et al. 2015) (compare also Section 4.”
Beyond Finite Layer Neural Networks: Bridging Deep Architectures and Numerical Differential Equations (2017, methodology): “Chen et al. (2015) proposed a feed-forward network in order to learn the optimal nonlinear anisotropic diffusion for image denoising.”
Unprocessing Images for Learned Raw Denoising (2018, neutral): “5%) TNRD [8] + VST 45.”
Conditioned Regression Models for Non-blind Single Image Super-Resolution (2015, background): “Still, modifying recent condition regression models like [5] would be an interesting research direction.”
Learning a Variational Network for Reconstruction of Accelerated MRI Data (2017, methodology): “Taking encouragement from early successes in image classiﬁcation tasks (3), recent advances also address semantic labeling (4), optical ﬂow (5) and image restoration (6).”
One Network to Solve Them All — Solving Linear Inverse Problems Using Deep Projection Models (2017, methodology): “[48] use a modified multi-stage non-linear diffusion process [11] to learn the proximal operator.”
Learning a Single Convolutional Super-Resolution Network for Multiple Degradations (2017, methodology): “Other CNN-based SISR methods can be found in [6, 22, 23, 30, 37, 42, 44, 45, 46, 53, 58].”
Fast Image Processing with Fully-Convolutional Networks (2017, background): “Other work has examined the acceleration of variational methods [6, 62, 13, 17], gradient-domain techniques [46], convolutions with large spatial support [23], and local Laplacian filters [5].”
Modern regularization methods for inverse problems (2018, methodology): “These cover the finitedimensional Markov random field models proposed in [325, 346, 143, 124, 334], the optimal model design approaches in [199, 198, 65, 40], the optimal regularization parameter estimation in variational regularization [89, 128, 137, 138, 90, 127], to training optimal operators in ”
Recurrent Inference Machines for Solving Inverse Problems (2017, methodology): “ents. Training an RIM from a practitioners perspective then boils down to choosing an RNN architecture. Because RNNs are Turing complete [1, 2], RIMs are a generalization of the models in Chen et al. [10], Gregor and LeCun [11], Wang et al. [12], Zheng et al. [13]. RIMs are also related to a recent ”
Bilevel approaches for learning of variational imaging models (2015, methodology): “Recently parameter learning in the context of functional variational regularisation models (1) also entered the image processing community with works by the authors [31, 18, 33, 32, 19, 25], Kunisch, Pock and co-workers [54, 22, 24], Chung et al. [27], and others [4, 74].”
Multi-channel Weighted Nuclear Norm Minimization for Real Color Image Denoising (2017, methodology): “We first compare MC-WNNM with the competing denoising methods [9, 11, 12, 14, 25, 28] on the 24 color images from the Kodak PhotoCD Dataset (http://r0k.”