Way to distinguish two different textures without dataset

asked 2020-09-02 05:15:06 -0600

I have video of moving agriculture harvester

image description

crop/grass can be different in all cases, but there is only two of them: cut and uncut. So i can't collect data set to train own neural network. -just don't have so many data or time to collect all possible cases.

Can i train own classifier on the fly, grabbing left and right roi? Or maybe somehow preprocess image. For example: using video optFlow i managed to receive something like this:

image description

however i can't achieve any result to separate cut/uncut areas with this grayscale image.

Any idea will be helpful

answered 2020-09-02 13:53:46 -0600

kbarni

4134 ●1 ●13 ●49

Well, this seems as a classical texture analysis problem.

There are some robust texture descriptors which can be used to differenciate between the two classes without prior training... or with a simple classifier, like SVM.

Check the Haralick descriptors on the image, Gabor filters, tensors, LBP (local binary pattern) etc. and see if there is a descriptor which can differenciate the cut/uncut textures.

P.S. the OptFlow estimates the movement speed of the different areas of the image. There is a difference between the (higher) uncut and cut parts. However this depends on the distance from the camera(distant parts move slower) and the distance from the center. It's also unreliable on zones with few caracteristic points (like uniform textures).

edit flag offensive delete link

Comments

Ok, this seems promising, thanks you! I think, i will try SVM approach

Samorodowv ( 2020-09-03 00:23:59 -0600 )edit

add a comment

Way to distinguish two different textures without dataset

1 answer

Comments

Links

Question Tools

Stats

Related questions

Way to distinguish two different textures without dataset edit

1 answer

Comments

Links

Question Tools

Stats

Related questions

Way to distinguish two different textures without dataset