Authors:
Yu Liu
1
;
Changwen Zheng
2
and
Hongliang Yuan
1
Affiliations:
1
Chinese Academy of Sciences and University of Chinese Academy of Sciences, China
;
2
Chinese Academy of Sciences, China
Keyword(s):
Adaptive Rendering, Image Space Reconstruction, Guided Image Filter, Mean Squared Error.
Related
Ontology
Subjects/Areas/Topics:
Applications
;
Computer Vision, Visualization and Computer Graphics
;
Geometry and Modeling
;
Image-Based Modeling
;
Image-Based Rendering
;
Pattern Recognition
;
Physics-Based Modeling
;
Rendering
;
Rendering Algorithms
;
Software Engineering
Abstract:
Image space rendering methods are efficient at removing Monte Carlo noise. However, a major challenge is
optimizing the bandwidth to denoise images while preserving their fine details. In this paper, a high-order
function is proposed to leverage the correlation between features and pixel colors. We consider feature buffers
to fit data while computing regression weights using pixel colors. A collaborative prefiltering framework is
first proposed to denoise features. The input pixel colors are then denoised using a guided image filter that
maintains fine details in the output by constructing a guidance image using features. The optimal bandwidth
is selected through an iterative error estimation process performed at multiple pixels to smooth the details.
Finally, we adaptively select center pixels to build our regression models and vary the window size to reduce
computational overhead. Experimental results showed that the new approach outperforms competing methods
in terms of the qualit
y of the visual image and the numerical error incurred.
(More)