This paper outlines the problem of noise in RGB-D images and proposes a solution that removes much of the noise in almost realtime (32ms for a 752x520 image).
Their process consists of 3 main phases.
- Use edge detector to identify salient edges across the image
- Apply jointly bilateral filter to the image to smoothen those regions with concinnous texture values while skip processing the parts with distinct structures
- Use the exemplar based scheme to find an optimal patch from those available depth region to the target region
Detailed Explanation of Each Step
Use of Edge Detection Algorithm
- Apply available edge detection algorithms such as Canny Edge Detection
- Detailed explanation of Canny Edge available here: https://en.wikipedia.org/wiki/Canny_edge_detector
- Implemented version of Canny Edge available here:
Apply jointly bilateral filter
- Application of jointly bilateral filter allows us to smooth parts of the image with concinnous texture values while skips processing the parts with distinct structures
- Specifically we use the given formula below
– Can’t Display due to 1 image restriction
where p and q represent the center location on image I for the
Gaussian Kernels f and g. kp is the normalizing factor; Ω is
the spatial range
Use of exemplar based scheme
- Finds an optimal patch from those available depth
region to the target region
- Adopt the strategy as Criminisi proposed on isophote-driven sampling process
– Can’t display due to 1 image restriction
where the metrics C( p ) and D( p ) are the confidence term and
data term for the priority patch definition P( p ) = C( p )·D( p ).