Denoising camera data - Emerging Trends in Image Processing, Computer Vision, and Pattern Recognition - page 6

(4)

This scale defines the shape dimension in the direction θ k . The ICI rule is applied in all eight

directions and thereby a neighborhood for the pixel position x is found. This neighbor-

hood can now be used for denoising.

3.3 Shape-adaptive DCT and Denoising via Hard Thresholding
The shape of the neighborhood is found for each pixel. Now the SA-DCT must be applied on
the Bayer data to perform the denoising. Therefore the Bayer data is separated into the four
sub-channels, R , G 1 , G 2 , and B , which each contain fourth of the total number of pixels.
For each sub-channel the SA-DCT is implemented as proposed by Foi [ 7 ] . A local estimate

is obtained by thresholding in the SA-DCT domain with the threshold parameter t x set to

(5)

The constant k thr regulates the denoising strength. The global estimate is given by a weighted
average of the local estimates.

(6)

The weights w x are calculated based on the size of the neighborhood and the noise variance

σ 2 . A smaller neighborhood gets higher weights, thus fine details are preserved. is the
number of non-zero coefficients after thresholding, thus sparse solutions are preferred over
non-sparse solutions.