viqa.fr_metrics.ssim.structural_similarity

viqa.fr_metrics.ssim.structural_similarity(img_r, img_m, win_size=None, data_range=None, gaussian_weights=True, alpha=1, beta=1, gamma=1, **kwargs)[source]

Compute the structural similarity index between two images.

Parameters:

  • img_r (np.ndarray) – Reference image to calculate score against.

  • img_m (np.ndarray) – Modified image to calculate score of.

  • win_size (int or None, optional) – The side-length of the sliding window used in comparison. Must be an odd value. If gaussian_weights is True, this is ignored and the window size will depend on sigma.

  • data_range (int, default=None) – Data range of the input images.

  • gaussian_weights (bool, default=True) – If True, each patch has its mean and variance spatially weighted by a normalized Gaussian kernel of width sigma=1.5.

  • alpha (float, default=1) – Weight of the luminance comparison. Should be alpha >=1.

  • beta (float, default=1) – Weight of the contrast comparison. Should be beta >=1.

  • gamma (float, default=1) – Weight of the structure comparison. Should be gamma >=1.

  • K1 (float, default=0.01) – Algorithm parameter, K1 (small constant, see [1]).

  • K2 (float, default=0.03) – Algorithm parameter, K2 (small constant, see [1]).

  • sigma (float, default=1.5) – Standard deviation for the Gaussian when gaussian_weights is True.

  • mode (str, default='reflect') –

    Determines how the array borders are handled. ‘constant’, ‘edge’, ‘symmetric’, ‘reflect’ or ‘wrap’.

    See also

    See Scipy documentation for scipy.ndimage.gaussian_filter() or scipy.ndimage.uniform_filter() for more information on the modes.

  • cval (float, optional) – Value to fill past edges of input if mode is ‘constant’. Default is 0.

Returns:

ssim – The mean structural similarity index over the image.

Return type:

float

Raises:

ValueError – If K1, K2 or sigma are negative. If win_size exceeds image or is not an odd number.

Warns:

RuntimeWarning – If alpha, beta or gamma are not integers.

Notes

To match the implementation in [1], set gaussian_weights to True and sigma to 1.5. This code is adapted from skimage.metrics.structural_similarity() available under [2]. The metric would possibly result in a value of nan in specific cases. To avoid this, the function replaces nan values with 1.0 before computing the final score.

References