BestScalingParamsFinder automates the process of finding optimal parameters for image scaling kernels. It resizes the scaled input image to the reference input image dimensions using range of parameter combinations, and evaluates the quality of each result by computing the Structural Similarity Index (SSIM) between the rescaled image and the reference image.
Uses Chris Lomont SSIM implementation.
Input images are, scaled image that will be rescaled to reference image size, and the reference image.
Images need to be 1 channel greyscale sRGB.
Use -h or --help to print help about all options.
If you want to lock any of parameters set "-lo" and "-hi" to the same value and set "-i" to 0.0.
Example usage and output:
BestScalingParamsFinder.exe --ref-img ref.png --scld-img scld.png --kernel 9 --p1-lo -1.0 --p1-hi 0.0 --p1-i 0.1
R: 2.000000, B: 1.000000, P1: -1.000000, P2: 0.000000, SSIM: 0.982032554157652
R: 2.000000, B: 1.000000, P1: -0.900000, P2: 0.000000, SSIM: 0.986954348755108
R: 2.000000, B: 1.000000, P1: -0.800000, P2: 0.000000, SSIM: 0.990853317338708
R: 2.000000, B: 1.000000, P1: -0.700000, P2: 0.000000, SSIM: 0.993772010927650
R: 2.000000, B: 1.000000, P1: -0.600000, P2: 0.000000, SSIM: 0.995797642013453
R: 2.000000, B: 1.000000, P1: -0.500000, P2: 0.000000, SSIM: 0.997040504634779
R: 2.000000, B: 1.000000, P1: -0.400000, P2: 0.000000, SSIM: 0.997646986076961
R: 2.000000, B: 1.000000, P1: -0.300000, P2: 0.000000, SSIM: 0.997776222153276
R: 2.000000, B: 1.000000, P1: -0.200000, P2: 0.000000, SSIM: 0.997583871415808
R: 2.000000, B: 1.000000, P1: -0.100000, P2: 0.000000, SSIM: 0.997218789477676
R: 2.000000, B: 1.000000, P1: 0.000000, P2: 0.000000, SSIM: 0.996792182037212
The best: R: 2.000000, B: 1.000000, P1: -0.300000, P2: 0.000000, SSIM: 0.997776222153276
Power of Cosine
n = p1
Has to be satisfied: n >= 0
n = 0: Box window
n = 1: Cosine window
n = 2: Hann window
Kaiser
beta = p1
Power of Garamond
n = p1, m = p2
Has to be satisfied: n >= 0, m >= 0
n = 0: Box window
m = 0: Box window
n -> inf, m <= 1: Box window
m = 1: Garamond window
n = 1, m = 1: Linear window
n = 2, m = 1: Welch window
Power of Blackman
a = p1, n = p2
Has to be satisfied: n >= 0
Has to be satisfied: if n != 1, a <= 0.16
n = 0: Box window
n = 1: Blackman window
a = 0, n = 1: Hann window
a = 0, n = 0.5: Cosine window
Generalized Normal Window (GNW)
s = p1, n = p2
Has to be satisfied: s != 0, n >= 0
n -> inf: Box window
n = 2: Gaussian window
Said
eta = p1, chi = p2
Has to be satisfied: eta != 2
Bicubic
a = p1
Has to be satisfied: radius = 2
Blur is unused.
Modified FSR Kernel
b = p1, c = p2
Has to be satisfied: radius = 2, b != 0, b != 2, c != 0
Blur is unused.
c = 1: FSR kernel
BC-Spline
b = p1, c = p2
Has to be satisfied: radius = 2
Blur is unused.
Keys kernels: B + 2C = 1
B = 1, C = 0: Spline kernel
B = 0, C = 0: Hermite kernel
B = 0, C = 0.5: Catmull-Rom kernel
B = 1 / 3, C = 1 / 3: Mitchell kernel
B = 12 / (19 + 9 * sqrt(2)), C = 113 / (58 + 216 * sqrt(2)): Robidoux kernel
B = 6 / (13 + 7 * sqrt(2)), C = 7 / (2 + 12 * sqrt(2)): RobidouxSharp kernel
B = (9 - 3 * sqrt(2)) / 7, C = 0.1601886205085204: RobidouxSoft kernel
You can use Visual Studio.
No additional dependencies.