A Novel Convex Autoregressive Model for Light Field Denoising on Riemannian Space
Mansi Sharma, Rohan Lal
In Conference Proceedings of European Association for Signal Processing (EURASIP),
European Light Field Imaging Workshop, Borovets, Bulgaria (ELFI 2019),
European Training Network on Full Parallax Imaging (Leading European Scientific Event)
Abstract
Existing light field cameras are susceptible to produce low-quality results as sensor noise can dominate measurements. Thus, denoising is a critical step in the light field (LF) subsequent analysis and processing. This paper presents a novel LF denoising framework based on an adaptive parallax and auto-regressive model analysis. The novel procedure first creates a view-dependent LF stack by compensating parallax variation employing an extended variational flow technique on a set of LF intensity and depth features. Further, it takes advantage of the spatial similarity across the registered LF stack and reduce the noisy observations. The output is, further, improved by formulating the denoising as a novel adaptive autoregressive (AR) stochastic problem. The proposed convex AR model averaged view-specific spatial energies of stacked LF images on Riemannian manifolds by a depth-directed maximum likelihood AR parameter estimation process. Lastly, scale intensity of refined AR LF predicted view by the average intensity of the superpixel in each LF stacked image. The experiments show that proposed AR LF denoiser outperforms standard algorithms in terms of visual quality and in the preservation of parallax details.
ELFI2019LFDenoiseCRpaper.pdfELFI2019PosterLFDenoise.pdfCitation
Mansi Sharma and Rohan Lal, A Novel Convex Autoregressive Model for Light Field Denoising on Riemannian Space, In European
Association for Signal Processing (EURASIP) Proceedings, European Light Field Imaging Workshop, Borovets, Bulgaria, ELFI2019.