Deep Gaussian Conditional Random Field Network: A Model-based Deep Network for Discriminative Denoising
Abstract: We propose a novel end-to-end trainable deep network architecture for image denoising based on a Gaussian Conditional Random Field (GCRF) model. In contrast to the existing discriminative denoising methods that train a separate model for each individual noise level, the proposed deep network explicitly models the input noise variance and hence is capable of handling a range of noise levels. Our deep network, which we refer to as deep GCRF network, consists of two sub-networks: (i) a parameter generation network that generates the pairwise potential parameters based on the noisy input image, and (ii) an inference network whose layers perform the computations involved in an iterative GCRF inference procedure. We train two deep GCRF networks (each network operates over a range of noise levels: one for low input noise levels and one for high input noise levels) discriminatively by maximizing the peak signal-to-noise ratio measure. Experiments on Berkeley segmentation and PASCALVOC datasets show that the proposed approach produces results on par with the state-of-the-art without training a separate network for each individual noise level.
Contributions
- We propose a new end-to-end trainable deep network architecture for image denoising based on a GCRF model. In contrast to the existing discriminative denoising methods that train a separate model for each individual noise level, the proposed network explicitly models the input noise variance and hence is capable of handling a range of noise levels.
- We propose a differentiable parameter generation network that generates the GCRF pairwise potential parameters based on the noisy input image.
- We unroll a half quadratic splitting-based iterative GCRF inference procedure into a deep network and train it jointly with our parameter generation network.
- We show that the proposed approach produces results on par with the state-of-the-art without training a separate network for each individual noise level.
Experimental results
Noise standard deviation | 10 | 15 | 20 | 25 | 30 | 35 | 40 | 45 | 50 | 55 | 60 |
Clustering SR | 33.27 | 30.97 | 29.41 | 28.22 | 27.25 | 26.30 | 25.56 | 24.89 | 24.28 | 23.72 | 23.21 |
EPLL | 33.32 | 31.06 | 29.52 | 28.34 | 27.36 | 26.52 | 25.76 | 25.08 | 24.44 | 23.84 | 23.27 |
BM3D | 33.38 | 31.09 | 29.53 | 28.36 | 27.42 | 26.64 | 25.92 | 25.19 | 24.63 | 24.11 | 23.62 |
NL-Bayes | 33.46 | 31.11 | 29.63 | 28.41 | 27.42 | 26.57 | 25.76 | 25.05 | 24.39 | 23.77 | 23.18 |
NCSR | 33.45 | 31.20 | 29.56 | 28.39 | 27.45 | 26.32 | 25.59 | 24.94 | 24.35 | 23.85 | 23.38 |
WNNM | 33.57 | 31.28 | 29.70 | 28.50 | 27.51 | 26.67 | 25.92 | 25.22 | 24.60 | 24.01 | 23.45 |
CSF | - | - | - | 28.43 | - | - | - | - | - | - | - |
MLP | 33.43 | - | - | 28.68 | - | 27.13 | - | - | 25.33 | - | - |
Proposed | 33.56 | 31.35 | 29.84 | 28.67 | 27.80 | 27.08 | 26.44 | 25.88 | 25.38 | 24.90 | 24.45 |
Training and test data
Use the below link to download the data used in our experiments.
Please cite the below paper if you use this data for your research.
Publications:
Raviteja Vemulapalli, Oncel Tuzel, and Ming-Yu Liu, "Deep Gaussian Conditional Random Field Network: A Model-based Deep Network for Discriminative Denoising", CVPR, 2016.
[PDF][Supplementary]
[PDF][Supplementary]
Patents:
Oncel Tuzel, Ming-Yu Liu and Raviteja Vemulapalli "Method and System for Denoising Images Using Deep Gaussian Conditional Random Field Network", US Patent 9,633,274, issued April 25, 2017.