Lepidoptera recognition identifier

Community and Forum → Website news and updates → Lepidoptera recognition identifier

Peter Khramov, 09.09.2014 22:54

First timid steps for fly to the moon.
So the first phase starts with spread specimens and, say, upperside angle shots. The recognition identifier should compare different shots of lepidoptera and ascertain quite well (more or less) if specimens are alike and how much.
For the beginning a set of photos needed to train this software and check its efficiency.
E.g., let's take a species and a few pics of spread specimens upperside view. Then take a couple of species alike and some corresponding images. Then less alike. And absolutely different. In the end we get a dozen or two of various images, different and alike. Then we can start training the software.
Sure there are some species completely identic in appearance, or subspecies which are more different from nominative species than totally different species, and so on. Obviously no magic here. Anyway, this is at least worth trying.
Any thoughts on that, what species to take for a start and pics of which species, different and alike, to use?

All the rest posts on: for your information, new features

Comments

10.09.2014 14:40, Peter Khramov

If the website content can't be used, no sense to try :--) it's not about the number of species. Alexandr, you think bookish, if there's no data enough then we get a limited selection of species (based on location or taxon or whatever else) which is not necessary here. Think wider, guys! The software should learn to determine the coefficient of the likeness between specimens in different pics (first in similar angle shots). Lots of many things could be got of this sole thing!
1. The software will be determine likeness between images on the website. The more images we got, the better the result.
2. The software will show specimens most alike or different. Or any other else after it calculates all the coefficients.
3. Results will be joined to catalogue selections. E.g., we add a pic and note that was Nymphalidae shot in Spain in July. No need to limit yourselves with only one feature.
That will be awesome. But only if coefficients will be calculated more or less adequately for which we need this very first set of images.
Once again: guys, there are really few limits if the website's well-organized (as against books). "Recognition identifier" doesn't mean at all it admits pics only and doesn't handle any other additional info.

10.09.2014 18:05, Shamil Murtazin

Peter!
First and foremost, the question arises, what are we, the neural network will build and teach her? Probably not.
Then, I think, the most important features of any mark which must be determined. If the configuration of the wings, venation (by the way, for the program may be easier to explain it than a man) and the pattern of the wing.I was just out of ignorance of the algorithm on which it's planned.
By type of - I do not see much of a difference that it is to define.
As a rule acted very much the same ph no shipping - more than three copies straightened in a form not found. This wayCatocala nupta, but there are other similar species.I have one unfolding, even a photo section for the collection did. I can download if necessary.
I suggestSaturnia pavonia: a well distinguished by color pattern, is "something which" unbundling geography.