Identifying mushrooms with an ML-based algorithm is a fine idea if you properly design the application to leverage that. As a hedgehog, this is what I would do:
Train my model on a variety of mushrooms, particularly poisonous ones.
When testing the model, test as many mushrooms as possible and take note of what’s frequently mis-identified.
When testing the model, make sure to get a variety of different kinds of lighting.
In addition to the mis-identifications noted while testing the app, maintain a list of commonly misidentified mushrooms - like the hedgehog mushroom and its counterparts - particularly the ones a forager should be most concerned with (meaning the most poisonous ones).
When identifying a mushroom to the user, err on the side of calling it a poisonous mushroom. Consider providing a list of possible matches, with the worst case scenario ones up top.
Include pictures and other information about the mushrooms, as well as regional mushroom lookups for mushrooms that weren’t included.
Don’t include text like “99% confident that this is a hedgehog mushroom” when the 99% figure is an output from your ML model. I know we said earlier to make sure to do a ton of testing and I’m sure you think you did, but you didn’t do enough to be able to say that. At best, reduce your certainty by 25%, then divide that number between the identified mushroom and the lookalikes, making sure to give extra weight to the most poisonous ones. So that 99% certainty becomes at most a more realistic 38% chance that it’s the poisonous lookalike and 37% chance that it’s whatever was identified in the first place.
You might say that this app would be useless for determining if a mushroom is safe to eat, and I agree, but it’s also a better approach than any of the existing apps out there. If you need to use an app to determine if a wild mushroom is safe to eat then the answer is simple: it isn’t. C’mon, I’m a hedgehog and even I know that.