Perceptual Image Hash

Personne n'a encore pris cette issue.

Thanks and congrats on `pytest-mpl`. It's really lovely :+1:

Over on [SciTools/iris](https://github.com/SciTools/iris) we've been using a very similar framework to perform graphical testing, but rather than using RMS or SHA-256 hashes we've opted to use perceptual image hashes.

Historically, we found using RMS/SHA-256 quite fragile and too sensitive to change, and that sometimes involved quite an overhead on devs to manually understand whether graphical test failures were significant or not; typically, it really depended on the nature of the image change involved, and that generally took time and effort to figure out.

However, for the last few years now we've adopted perceptual image hashing as a more robust and stable approach to graphical image testing, and we'd love to be able to use `pytest-mpl` as a general framework (rather than reinvent the wheel) as we're keen to use perceptual image hashing elsewhere in other packages other than `iris`.

If you want to know more about perceptual image hashing:
- here's a [gist notebook](https://gist.github.com/bjlittle/6440844ea509276edf4bfbea18c8d41a) detailing outcomes from our initial investigation of perceptual image hashing as an option for `SciTools/iris`
- here's a [scientific paper](https://www.sciencedirect.com/science/article/pii/S1877050921011030) (of which there are many) performing an analysis of perceptual image hashing algorithms

In the end, we opted to leverage the https://github.com/JohannesBuchner/imagehash package, which is available on both `conda-forge` and `pypi`, to perform the actual perceptual image hashing.

At this point, I guess my question to the `pytest-mpl` core developers are:
1. Have you considered using perceptual image hashing as an approach and decided not to use it? If so, why? That would be really great to know.
2. If you're not against the concept, would you be open to us making an additive change to `pytest-mpl` where the user can choose to opt-in to an alternative hashing algorithm, that they configure, to perform their graphical testing?

I think that option [2] would be possible and a valuable feature for the `pytest-mpl` community.

However, you guys have the vision of where you want `pytest-mpl` go and the use cases that you want to address. It would be really great to know whether you're open to offering configurable hashing kernels as a feature. If you are, then we'd be more than willing to do the work and contribute to `pytest-mpl` under your guidance and advice.

Thanks :smiley:

Commencez par examiner le comportement existant de pytest-mpl en matière de comparaison d’images graphiques et de hachage, puis étudiez comment le paquet imagehash pourrait prendre en charge une alternative opt-in. Le travail sera considéré comme terminé lorsqu’une décision de projet aura été prise et, si elle est acceptée, qu’une fonctionnalité configurable de hachage perceptuel aura été définie, avec une couverture de son comportement de comparaison d’images.

Rédigé par le modèle d'indexation à partir du texte de l'issue.