BRISQUE#

class deepinv.loss.metric.BRISQUE(weights_path='download', max_pixel=1.0, device='cpu', dtype=torch.float32, **kwargs)[source]#

Bases: Metric

Blind/Referenceless Image Spatial QUality Evaluator (BRISQUE) metric.

Calculates the BRISQUE score \(\text{BRISQUE}(\hat{x})\) where \(\hat{x}=\inverse{y}\). It is a no-reference image quality metric introduced by Mittal et al.[1], which quantifies how far an image departs from the natural scene statistics of pristine natural images. Lower is better, with scores roughly in \([0, 100]\).

BRISQUE works with images of 1 or 3 channels. If the image has 3 channels, it is assumed to be RGB and converted to relative luminance, then, at two scales, the mean subtracted contrast normalized (MSCN) coefficients

\[\hat{x}_{ij} = \frac{x_{ij} - \mu_{ij}}{\sigma_{ij} + 1}\]

are computed with a \(7\times 7\) Gaussian window. A generalized Gaussian is fitted to the MSCN coefficients and asymmetric generalized Gaussians are fitted to their four neighbouring products, yielding 36 features which are mapped to a quality score by a support vector regressor pre-trained on the LIVE IQA dataset.

This is a PyTorch translation of the implementation (dsoellinger/blind_image_quality_toolbox). The pre-trained support vector regressor is the one released with the original MATLAB implementation, and weights were downloaded from dsoellinger/blind_image_quality_toolbox.

Note

The features and the regressor were fitted on images in the \([0, 255]\) range. Inputs are internally rescaled from [0, max_pixel] to \([0, 255]\), so make sure max_pixel matches the intensity scale of your data.

Note

By default, no reduction is performed in the batch dimension.

Parameters:

  • weights_path (str, pathlib.Path, None) – path to the support vector regressor weights. If 'download' (default), the weights released with the original implementation are downloaded.

  • max_pixel (float) – maximum pixel value of the input images, used to rescale them to the \([0, 255]\) range expected by the regressor. Default: 1.

  • device (torch.device, str) – device on which the regressor weights are stored. Default: 'cpu'.

  • dtype (torch.dtype) – dtype used for the feature computation (the regressor is always evaluated in float64). Default: torch.float32.

  • complex_abs (bool) – perform complex magnitude before passing data to metric function. If True, the data must either be of complex dtype or have size 2 in the channel dimension (usually the second dimension after batch).

  • reduction (str) – a method to reduce metric score over individual batch scores. mean: takes the mean, sum takes the sum, none or None no reduction will be applied (default).

  • norm_inputs (str) – normalize images before passing to metric. l2 normalizes by \(\ell_2\) spatial norm, min_max normalizes by min and max of each input.

  • center_crop (int, tuple[int], None) – If not None (default), center crop the tensor(s) before computing the metrics. If an int is provided, the cropping is applied equally on all spatial dimensions (by default, all dimensions except the first two). If tuple of int, cropping is performed over the last len(center_crop) dimensions. If positive values are provided, a standard center crop is applied. If negative (or zero) values are passed, cropping will be done by removing center_crop pixels from the borders (useful when tensors vary in size across the dataset).

Example:

>>> from deepinv.loss.metric import BRISQUE
>>> m = BRISQUE()
>>> x_net = torch.rand(2, 3, 32, 32)  # batch of 2 RGB images in [0, 1]
>>> m(x_net).shape
torch.Size([2])

References:

estimate_aggd_param(vecs, eps=1e-12)[source]#

Fit an asymmetric generalized Gaussian distribution to each row by moment matching.

Parameters:

  • vecs (torch.Tensor) – (B, N) tensor of samples.

  • eps (float) – stabilizer used in the denominators.

Returns:

tuple of (B,) tensors with the shape parameter, the left and the right standard deviations.

Return type:

tuple[Tensor, Tensor, Tensor]

estimate_ggd_param(vecs, eps=1e-12)[source]#

Fit a generalized Gaussian distribution to each row by moment matching.

Parameters:

  • vecs (torch.Tensor) – (B, N) tensor of samples.

  • eps (float) – stabilizer used in the denominators.

Returns:

tuple of (B,) tensors with the shape parameter and the standard deviation.

Return type:

tuple[Tensor, Tensor]

features(x_net)[source]#

Compute the 36 natural scene statistics features of BRISQUE.

Parameters:

x_net (torch.Tensor) – (B, 1, H, W) single-channel images in the [0, 255] range.

Returns:

(B, 36) tensor of features.

Return type:

Tensor

metric(x_net, *args, **kwargs)[source]#

Compute the BRISQUE score for a batch of images.

Parameters:

x_net (torch.Tensor) – (B, C, H, W) input tensors with C=1 or 3 channels.

Returns:

(B,) tensor of BRISQUE scores.

Return type:

Tensor