ssim#

diffsptk.SSIM#: alias of StructuralSimilarityIndex

class diffsptk.StructuralSimilarityIndex(reduction='mean', *, alpha=1, beta=1, gamma=1, kernel_size=11, sigma=1.5, k1=0.01, k2=0.03, eps=1e-08, padding='same', dynamic_range=None)[source]#

Structural similarity index computation.

Parameters:

reduction[‘none’, ‘mean’, ‘sum’]: Reduction type.
alphafloat > 0: Relative importance of luminance component.
betafloat > 0: Relative importance of contrast component.
gammafloat > 0: Relative importance of structure component.
kernel_sizeint >= 1: Kernel size of Gaussian filter.
sigmafloat > 0: Standard deviation of Gaussian filter.
k1float > 0: A small constant.
k2float > 0: A small constant.
epsfloat >= 0: A small value to prevent NaN.
padding[‘valid’, ‘same’]: Padding type.
dynamic_rangefloat > 0 or None: Dynamic range of input. If None, input is automatically normalized.

References

[1] Z. Wang et al., “Image quality assessment: From error visibility to structural: similarity,” IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004.

forward(x, y)[source]#

Calculate SSIM.

Parameters:

xTensor [shape=(…, N, D)]: Input.
yTensor [shape=(…, N, D)]: Target.

Returns:

outTensor [shape=(…, N, D) or scalar]: SSIM or mean SSIM.

Examples

>>> x = diffsptk.nrand(20, 20)
>>> y = diffsptk.nrand(20, 20)
>>> ssim = diffsptk.StructuralSimilarityIndex()
>>> s = ssim(x, y)
>>> s
tensor(0.0588)

diffsptk.functional.ssim(x, y, reduction='mean', *, alpha=1, beta=1, gamma=1, kernel_size=11, sigma=1.5, k1=0.01, k2=0.03, eps=1e-08, padding='same', dynamic_range=None)[source]#

Calculate SSIM.

Parameters:

xTensor [shape=(…, N, D)]: Input.
yTensor [shape=(…, N, D)]: Target.
reduction[‘none’, ‘mean’, ‘sum’]: Reduction type.
alphafloat > 0: Relative importance of luminance component.
betafloat > 0: Relative importance of contrast component.
gammafloat > 0: Relative importance of structure component.
kernel_sizeint >= 1: Kernel size of Gaussian filter.
sigmafloat > 0: Standard deviation of Gaussian filter.
k1float > 0: A small constant.
k2float > 0: A small constant.
epsfloat >= 0: A small value to prevent NaN.
padding[‘valid’, ‘same’]: Padding type.
dynamic_rangefloat > 0 or None: Dynamic range of input. If None, input is automatically normalized.

Returns:

outTensor [shape=(…, N, D) or scalar]: SSIM or mean SSIM.