There is a set of Pixel Level transforms that is used in the work Benchmarking Neural Network Robustness to Common Corruptions and Perturbations

The authors also share the code => we can absorb some transforms that they have into the library.

https://github.com/hendrycks/robustness/blob/master/ImageNet-C/create_c/make_imagenet_c.py

• ShotNoise
• Defocus
• GlassBlur

Initial attempt failed, we need to fix the following issues:
SixLabors/ImageSharp#1302 (comment)

Disabling the two (likely) bugs with [ActiveIssue] might be also fine.

Description

While reviewing #5006, I noticed that threshold_li repeatedly computes the mean pixel value above threshold and the mean below threshold. For integer images, this could be done with the histogram rather than with the full image, using a weighted mean. Possibly the approximation with float images would be good enough for gradient descent anyway.

See https://scikit-image.org/