tfmri.losses.ssim_multiscale_loss

ssim_multiscale_loss(y_true, y_pred, max_val=None, power_factors=(0.0448, 0.2856, 0.3001, 0.2363, 0.1333), filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03, batch_dims=None, image_dims=None, rank=None)

Computes the multiscale structural similarity (MS-SSIM) loss. (deprecated arguments)

Deprecated: SOME ARGUMENTS ARE DEPRECATED: (rank). They will be removed after 2022-09-01. Instructions for updating: Use argument image_dims instead.

The MS-SSIM loss is equal to \(1.0 - \textrm{MS-SSIM}\).

Parameters

  • y_true – A Tensor. Ground truth images. For 2D images, must have rank >= 3 with shape batch_shape + [height, width, channels]. For 3D images, must have rank >= 4 with shape batch_shape + [depth, height, width, channels]. height, width and depth must be greater than or equal to (filter_size - 1) * 2 ** (len(power_factors) - 1) + 1. Must have floating point type, with values in the range [0, max_val].

  • y_pred – A Tensor. Predicted images. For 2D images, must have rank >= 3 with shape batch_shape + [height, width, channels]. For 3D images, must have rank >= 4 with shape batch_shape + [depth, height, width, channels]. height, width and depth must be greater than or equal to (filter_size - 1) * 2 ** (len(power_factors) - 1) + 1. Must have floating point type, with values in the range [0, max_val].

  • max_val – The dynamic range of the images (i.e., the difference between the maximum and the minimum allowed values). Defaults to 1 for floating point input images and MAX for integer input images, where MAX is the largest positive representable number for the data type.

  • power_factors – A list of weights for each of the scales. The length of the list determines the number of scales. Index 0 is the unscaled resolution’s weight and each increasing scale corresponds to the image being downsampled by 2. Defaults to (0.0448, 0.2856, 0.3001, 0.2363, 0.1333), which are the values obtained in the original paper.

  • filter_size – The size of the Gaussian filter. Defaults to 11.

  • filter_sigma – The standard deviation of the Gaussian filter. Defaults to 1.5.

  • k1 – Factor used to calculate the regularization constant for the luminance term, as C1 = (k1 * max_val) ** 2. Defaults to 0.01.

  • k2 – Factor used to calculate the regularization constant for the contrast term, as C2 = (k2 * max_val) ** 2. Defaults to 0.03.

  • batch_dims –

    An int. The number of batch dimensions in input images. If None, it is inferred from inputs and image_dims as (rank of inputs) - image_dims - 1. If image_dims is also None, then batch_dims defaults to 1. batch_dims can always be inferred if image_dims was specified, so you only need to provide one of the two.

  • image_dims –

    An int. The number of spatial dimensions in input images. If None, it is inferred from inputs and batch_dims as (rank of inputs) - batch_dims - 1. Defaults to None. image_dims can always be inferred if batch_dims was specified, so you only need to provide one of the two.

  • rank –

    An int. The number of spatial dimensions. Must be 2 or 3. Defaults to tf.rank(y_true) - 2. In other words, if rank is not explicitly set, y_true and y_pred should have shape [batch, height, width, channels] if processing 2D images or [batch, depth, height, width, channels] if processing 3D images.

Returns

A Tensor of type float32 and shape batch_shape containing an SSIM value for each image in the batch.

References

1

Zhao, H., Gallo, O., Frosio, I., & Kautz, J. (2016). Loss functions for image restoration with neural networks. IEEE Transactions on computational imaging, 3(1), 47-57.