[R-sig-eco] Questions about pattern recognition of weed photos

[Jeremy Chacon]

It's possible, but I don't think it would be easy.

Also, I wouldn't write it in R, but that's only because I don't know how to
interface R with a phone. Maybe someone else has ideas.

What I would do is first download all the images onto your own computer,
and write a supervised machine learning algorithm to learn the features of
the identified images. Hopefully there are many images per weed, or else
this will never work!

I would use the openCV library, in my experience the java or python
bindings are reasonably easy to work with.

Once I was happy with the model (not a small task), I'd write an android
app (in java) that also uses opencv to pull out the relevant features from
the image you took, and runs those features through the model you
previously made, for prediction.

You'd probably also want to show the user what the model thinks is the
right weed, rather than just telling its name. You could do this by having
all the images on your app, but I think it'd be more efficient to hold them
on a server somewhere and call up images as needed. You could get even
fancier then and if the user agreed with the identification, save that set
of features to improve the model. If the use disagreed, save that too, and
maybe show them the next-best match.

Really this is a huge task and I wouldn't attempt it unless you have half a
year of dedicated time for it. But maybe some experts will have useful
shortcuts I don't know about. Good luck!

On Thu, Aug 27, 2015 at 3:07 PM, Richard Rice <rrice899 at gmail.com> wrote:

>  First some background. I
> collect weed species data in the field for the Washington state dept
> of natural resources(DNR). There are several Android apps for
> identifying weeds by looking at photos but none that I can find by
> comparing a photo that I take with my Android phone with a
> downloadable database of known weed species. What I'd like to do is
> write an app that downloads DNR's weed photos of which there are about
> 1200 to my phone, find a weed in the field, take a photo and compare
> it with DNR's photos to see if there is a match. This would save me a
> lot of time and hopefully reduce errors.
> Does this sound feasible?
> Is R capable of this?
>
> Best Regards,
>
> Richard
>
> _______________________________________________
> R-sig-ecology mailing list
> R-sig-ecology at r-project.org
> https://stat.ethz.ch/mailman/listinfo/r-sig-ecology
>



-- 

*___________________________________________________________________________Jeremy
M. Chacon, Ph.D.*

*Post-Doctoral Associate, Harcombe Lab*
*University of Minnesota*
*Ecology, Evolution and Behavior*

	[[alternative HTML version deleted]]