Weighted multi-band cross color channel predictor

1. A method for inter-layer prediction using a computer, the method comprising:

accessing a first image ( 145 ) and a second image ( 125 ), each of the images comprising one or more color channels, wherein both images represent the same scene but at a different dynamic range and the second image ( 125 ) has a higher dynamic range than the first image ( 145 );

for at least one color channel of the first image:

selecting a multivariate locally weighted (MLW) prediction model using two or more bands;

for each of the two or more bands of the selected MLW model:

computing prediction weights for the band;

solving for prediction parameters for the band and computing a band-based predicted output based on the prediction weights of the band, the first image, the second image, and the MLW prediction model ( 520 ); and

generating a final predicted output image based on the prediction weights of all the bands and the band-based predicted outputs ( 530 , 540 ).

2. The method of claim 1 , wherein the prediction weights are based on one or more exponential functions of pixel values in the first image.

3. The methods of claims 2 , wherein the prediction weights for identical pixel values for all bands add to one.

4. The method of claim 2 , wherein the prediction weights further comprise a positive decay parameter.

5. The method of claim 4 , wherein the decay parameter is adapted according to an adaptation function.

6. The method of claim 5 wherein the adaptation function for the decay parameter comprises a function of clipped pixels in the first image.

7. The method of claim 1 , further comprising:

compressing the first image into an encoded base layer bitstream; and

compressing an image derived from the second image and the prediction output image into one or more encoded enhancement layer bitstreams.

8. The method of claim 1 wherein the first image comprises a standard dynamic range (SDR) image and the second image comprises a visual dynamic range (VDR) image.

9. The method of claim 4 , wherein the prediction parameters of the MLW model and the decay parameter are transmitted to a decoder as metadata.

10. The method of claim 2 , wherein for a two-band weighted prediction MLW model the prediction weight functions comprise:

w

i

,

c

u

=

0

=

exp

(

-

(

1

-

d

i

)

2

h

)

for a first band and w i,c u=1 =1−w i,c u=0 for a second band, wherein h denotes a decay parameter and d i is based on an input image pixel value.

11. The method of claim 2 , wherein for a three-band weighted prediction MLW model the prediction weight functions comprise:

w

i

,

c

u

=

0

=

exp

(

-

(

d

i

j

=

0

)

2

h

u

=

0

)

for a first band,

w

i

,

c

u

=

1

=

exp

(

-

(

d

i

j

=

1

)

2

h

u

=

1

)

for a second band, and w i,c j=2 =1−w i,c u=0 −w i,c u=1 for a third band, wherein h u=0 and h u=1 denote decay parameters for the first and the second band, and d i j=0 and d i j=1 are based on input image pixel values.

12. The method of claim 6 wherein the adaptation function comprises

h

=

α

⁢

N

P

+

h

0

wherein h 0 denotes an initial decay parameter, P denotes the total number of pixels in the first image, α denotes a scaling factor, and N denotes a number of clipped pixels in the first image.

13. The method of claim 12 , wherein a pixel is counted as clipped when at least one of its color components is larger than a first threshold and a corresponding color component of a co-located pixel in the second image is smaller than the first threshold.

14. The method of claim 12 , wherein a pixel is counted as clipped when at least one of its color components is lower than a second threshold and a corresponding color component of a co-located pixel in the second image is larger than the second threshold.

15. The method of claim 1 , wherein generating a final predicted output image based on the prediction weights and the band-based predicted outputs comprises computing

{circumflex over (v)} c =Σ u=0 U-1 W c u {circumflex over (v)} c u ,

wherein W c u denotes prediction weights for an u-th band and {circumflex over (v)} c u denotes a band-based predicted output for the u-th band.

16. The method of claim 15 , wherein computing a band-based predicted output {circumflex over (v)} c u comprises

{circumflex over (v)} c u =Sm c u ,

wherein S denotes a matrix of pixel data in the first image based on the prediction model and m c u denotes a vector of prediction coefficients for the u-th band.

17. The method of claim 16 , wherein computing the prediction coefficients for the u-th band comprises computing

m c u =( S T W c u S ) −1 ( S T W c u v c ),

wherein v c denotes a vector of pixel data based on the second image.

18. An image decoding method comprising:

accessing a first image ( 490 ) with a first dynamic range;

for at least one color channel of the first image:

accessing metadata comprising data for a multivariate locally weighted (MLW) prediction model, the metadata comprising prediction parameters for two or more bands;

generating prediction weights for the MLW prediction model based on the metadata;

generating band-based predicted outputs based on the prediction weights and the first image; and

generating a final output predicted image based on the prediction weights and the band-based predicted outputs.

19. The method of claim 18 , further comprising:

accessing a residual image;

combining the residual image and the final output predicted image to generate a second image, wherein the second image has a second dynamic range higher than the first dynamic range.

20. The method of claim 19 wherein the first image comprises an SDR image and the second image comprises a VDR image.

21. A non-transitory computer-readable storage medium having stored thereon computer-executable instruction for executing a method with a computer in accordance with claim 1 .

22. A non-transitory computer-readable storage medium having stored thereon computer-executable instruction for executing a method with a computer in accordance with claim 18 .