Using machine learning for fish identification on MFK

[Captain_Kiwi]

One of the most common posts I see on MFK is people asking to ID species, what if MFK had a way to automate this process You could just upload an image and you get the most likely species.

A bit of background - I am a computer science student who has experience training image recognition models (I've also taken multiple modules in the subject) and I'm currently looking for a dissertation project.

Theoretically if I was to do this I would need around 1000 photos per "class" i.e sub species and that is by far the biggest hurdle. Technically people could just leave there camera on all night taking photos every 10 seconds and over 8 hours that would 2280 images, but I'm not a camera guy and I'm not sure if that's a thing? I was just interested would the community be in something like this because it's definitely a cool idea.

 

[esoxlucius]

I'd say most, if not all of those identification posts are for cichlids, usually people trying to identify a young fish, where it's notoriously difficult anyway, especially if it's hybridised.

So for your ID idea to work you'd need definite slam dunk ID's on all those type of fish if it was going to be 100% successful.

And there lies the problem because most of the time, whilst the fish are young, those ID's are a pure crapshoot anyway.

So your system would be no better than the lottery we have now imo.

 

[Captain_Kiwi]

I'd say most, if not all of those identification posts are for cichlids, usually people trying to identify a young fish, where it's notoriously difficult anyway, especially if it's hybridised.

So for your ID idea to work you'd need definite slam dunk ID's on all those type of fish if it was going to be 100% successful.

And there lies the problem because most of the time, whilst the fish are young, those ID's are a pure crapshoot anyway.

So your system would be no better than the lottery we have now imo.

Not slam dunked IDs, just as many labeled photos as possible of confirmed species. Blurry photos work fine as long as there are also clear photos. Interestingly it could be good at recognising hybrids - when predicting it give the probabilities of what it thinks the class is, so as long as the fish has physical characteristics from both species it should theoretically be able to recognise that.