External prior learning and internal mean sparse coding for image denoising

Authors: LYU, Q - Shanxi University; GUO, M - Shanxi University; MA, M - Shanxi University; Mankin, Richard

Submitted to: Journal of Electronic Imaging
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/1/2019
Publication Date: 5/23/2019
Citation: Lyu, Q., Guo, M., Ma, M., Mankin, R.W. 2019. External prior learning and internal mean sparse coding for image denoising. Journal of Electronic Imaging. 28(3):033014. https://doi.org/10.1117/1.JEI.28.3.033014.
DOI: https://doi.org/10.1117/1.JEI.28.3.033014

Interpretive Summary: Denoising of electronic signals is an important problem in acoustic and image analysis. Scientists and students at the Shaanxi Normal University, in collaboration with a scientists at the USDA-ARS Center for Medical, Agricultural, and Veterinary Entomology, Gainesville, Florida, developed new signal processing methods to remove noise from images. The methods are generally applicable to a variety of situations where some of the signals are relatively noise-free but others contain noise that was present when the signals were being collected. In this analysis, images were denoised, but the same process can be used to denoise field recordings of hidden infestations of insects in trees or soil and more broadly to non-agricultural acoustic and image data.

Technical Abstract: Image prior and sparse representation play an important role in image denoising. Many denoising methods learn priors either from noisy images itself or an external clean image dataset. But, only using external priors or noisy image itself priors cannot reconstruct the image effectively. In addition, since the image is corrupted by noise, the local sparse coding coefficient which obtained from the noisy image patch is also not accurate enough. Hence, the denoising performance is greatly restricted. In this paper, we present a new image denoising framework based on external prior learning and internal mean sparse coding method (EPL-IMSC). Specifically, we obtain external priors from a clean natural image dataset by Gaussian Mixture Model (GMM). The external priors are exploited to guide the subspace clustering of internal noisy image patches, and a compact dictionary is generated for each internal noisy patch cluster. Then, the internal mean sparse coding strategy based on non-local self-similarity (NSS) is introduced into the sparse representation model, whose regularization parameters are deduced by Bayesian framework. An iterative shrinkage method is employed to solve l1-optimization problem of this sparse representation model. Experimental results on different test images show that the proposed model exceeds other comparison methods in denoising performance.