Hello

I think what you are looking for is not a solution for individual identification within a class (whether this is a specific person or a specific animal for behavioral studies). What you want is a solution for identification of different classes (i.e. supervised learning for classifying an insect to a specific species). 

In other words, (for the example of dogs) you dont want to determine if that dog is Rex or Max (differentiate between individuals of the same class) but you want to be able to say if that dog is a sheepdog or a golden retriever (identification of classes/spieces).

Same idea for insects. Right?

Ok so....

There are many solutions available for generic classes of objects or animals (a google search will give you many results and software) but I dont know any specific for insects. The generic methods should work and you just train them on insect data (whether these are images or physiological data or even sound!). However, perhaps a more specialised supervised learning technique that has been specifically tailored for insects might be better (i.e. using features that are more discriminative for insect speciation).

For example, a quick google search recovered this

http://www.bmva.org/bmvc/2010/conference/paper108/paper108.pdf

You can ask the authors for the code or look at relevant papers that reference this or are referred by this.

So in summary, yes I am sure people have done this type of work (e.g. reference above). It is just a matter of doing the search and finding out

keywords you might like to try

supervised learning, insect classification, learning insect features for classification,

Best of luck,

Vasileios

I suggest the methods on these publications, these methodologies are able to identify: fossil diatoms, phytoplankton, finger prints , alphabet letters.

Take a look and try with insects.

Regards

Chapter Pattern Recognition of Digital Images by One-Dimensional Signatures

Conference Paper The digital pattern recognition system by Fourier masks.

Article Adaptive nonlinear correlation with a binary mask invariant ...

I am very pesimistic to such possibility. It possible to identify Insects visualy to a high degree of reliability up the generic level ind in some species up to the species level. But it needs a longterm experience in identification of a group. But in insects in the large genera with several tens of species, like Amara, Philonthus, Quedius the distinguishing characters are to gentil and, at the same time variable. Consulting the key or collection material is necessary. In addition the characers are situated in different sides of body and often need a manipulation with the identified insects, obseration from different angles etc. It is not a simple structure based on distance of eyes and ip of nose and chin, like in human face.

In addition ability to identify a group must be  closely connected with a good knowledge of  ecology of individual species and field experinece in collecting representatives of the respective group in a variety of habitats. This is an inevitable precondition of a correct monitoring or bioindication. Interretation of such data is not a meditation aboev a list o names telling nothing to the respective person. The sophisticed statistic program can help to visualize different relationships, but the final interpretation must done by an experienced person, not by an enthusiast engaged ad hoc for some task

Vasileios, you're right. What I am searching for is a software to identify classes. In this case, the classes are butterfly species that occur in a defined area. The taxonomic identification should not be as difficult as put by Zbyšek, as we aim at butterflies, which are usually recognised to species level by color-shape pattern. Another imortant issue is that at any given site, the total number of species is somewhat limited, so possible misidentification by the software may be constrained by the locality of occurrence. For precise taxonomic identification a real life taxonomist will always be necessary. Taxonomists will also be needed to teach the software. The purpose of such software would be to help lay people to engage in biodiversity monitoring schemes in the tropical region, as a way to help them learn faster and gather better data.

Hello again,

yes I understand. I think specifically for butterflies (where shape and colour are the most important characteristics for classification) you should have much more luck finding machine learning solutions, than just for general insect species.

Again a quick google search returns many papers  So you have a much larger choice

e.g.

https://www2.ia-engineers.org/iciae/index.php/icisip/icisip2014/paper/view/385

http://cvmldm.avestia.com/2015/PDF/002.pdf

http://www.temjournal.com/documents/vol2no1/pdf/A%20computer%20vision%20system%20for%20the%20automatic%20identification%20of%20butterfly%20species%20via%20Gabor-filter-based%20texture%20features%20and%20extreme%20learning%20machine%20GF+ELM.pdf

http://www.tandfonline.com/doi/abs/10.1080/0952813X.2013.861875#.VOyXXjVVKlM

Since I am not an expert on using computer vision for butterfly classification I cannot recommend one specific paper. Your best bet is to look at the results of the different methods and make your selection based on accuracy and software availability.

Of course what Zbyšek said is right and you need expert knowledge for proper taxonomic classification. However, if you want a tool that will help you shift through lots of images to provide a good-enough basic classification then by all means computer vision/machine learning is the right tool.

As with face recognition (or any other application) they are not fire and forget tools. They will give you some good result but they (probably) will never surpass human/manual classification especially when it comes to expert fields.

As you said "tools to help people learn faster and gather better data". Thats how they should be used and I totally agree with that.

Best of luck,

Vasileios

I suggest you to make contact with Dr. Josue Alvarez Borrego, maybe he can help you. His email is [email protected]

Thanks, Vasileios for the references and Cristina for the indication.

Best,

Onildo

Hi Onildo, Your problem is actually the topic of my dissertation!  I'm creating a software framework (in Mathematica) that will allow others to design, train, and test a classifier--an algorithm that classifies individuals to some given taxonomic level (genus, species, maybe even individual, depending on the dataset)--for their organism of interest.  I'm optimizing a species classifier for dragonfly and damselfly (based on scans of their wings), which has achieved 85-90% accuracy in classifying 14 odonate species.  However, I'm trying to build the framework generically enough that it could be used for a wider range of insects or other taxa, with minimal coding required.  To classify individuals based on their wing scan, I'm pulling information related to wing patterning (with Gabor wavelet transformation) and color.  In addition, I'm also borrowing algorithms from facial recognition software to create an automatic landmarking system from which additional identifying features could be pulled from individual images.  I still have a ways to go on this last part, but I think it has the potential to greatly increase the accuracy of the classifiers.

I think this system could be well-suited to your butterflies, assuming that you could get more-or-less standardized photos (or scans) of them.  I hope to finish everything in the next year or so and would like to have another dataset on which to test the system.  I'd love to hear more about the project that you've got in mind.  Please email me ([email protected]) if you'd like to talk more about this.

Cheers,

Will 

Will,

It is great to hear about your work. It fits perfectly in what we are trying to do here in Brazil. It would be great to test your algorithm with butterflies. The standardization of the photos may be an issue we will have to work. As to be of any help to a citizen scientist, the app/software must be able to recognize incomplete photos (body partially covered by fingers) and lateral images (ventral side of one of the wings), which are easier to take in the field when the insects are perched.

We can discuss more about that by e-mail.

Thanks a lot for your help!

Hey Onildo,

I agree that to be useful, automatic classification software must be versatile and flexible with respect to image quality and standardization.  Right now, my system only works with images taken under very strict conditions, but I'd like to make it more general, and therefore useful for others.  Your images could help me to do that!  I eagerly await your email (to [email protected]) with more details about the data that you are working with and the images that you have (e.g. how many species do you have? are they all butterflies? how many images per species? under what conditions are the images taken?).

Cheers,

Will