UniversalImageQualityIndex & SpectralDistortionIndex are still possible to give NaN results

[michael080808]

🐛 Bug

I noticed the issue #1520.
The fix may still cause "divided by zero" problem because eps is not directly added to the divisor.

To Reproduce

    upper = 2 * sigma_pred_target
    lower = sigma_pred_sq + sigma_target_sq + torch.finfo(sigma_pred_sq.dtype).eps

    uqi_idx = ((2 * mu_pred_target) * upper) / ((mu_pred_sq + mu_target_sq) * lower)

Here (mu_pred_sq + mu_target_sq) * lower still may contain zero values. For example, if there are several pictures with large black area, both mu_pred_sq and mu_target_sq can be zeros. To reproduce the problem, these two images can be as examples.

Expected behavior

    upper = 2 * sigma_pred_target
    lower = sigma_pred_sq + sigma_target_sq

    uqi_idx = ((2 * mu_pred_target) * upper) / ((mu_pred_sq + mu_target_sq) * lower + torch.finfo(lower.dtype).eps)

This fix is very easy to understand. Just thanks to your excellent jobs!

Environment

Name	Version	Build	Channel
_anaconda_depends	2023.07	py311_0	https://repo.anaconda.com/pkgs/main
conda	23.9.0	py311haa95532_0	https://repo.anaconda.com/pkgs/main
python	3.10.13	he1021f5_0	defaults
pytorch	2.1.0	py3.10_cuda12.1_cudnn8_0	pytorch
pytorch-cuda	12.1	hde6ce7c_5	pytorch
pytorch-mutex	1.0	cuda	pytorch
torchaudio	2.1.0	pypi_0	pypi
torchmetrics	1.2.0	pyhd8ed1ab_0	conda-forge
torchvision	0.16.0	pypi_0	pypi

[github-actions]

Hi! thanks for your contribution!, great first issue!