1

I have an image of pellets (balls). I need to find part of image, where balls are.

My testing image looks like this: Example of image

I have an idea using wavelets. In 1D, ball could be quite well recognized by Continuou wavelet transform.

Here is example 2D signal of ball: Ball signal

Here is Continuous wavelet transform of this signal using Morlet wavelet: CWT of signal

And here is just very simple threshold of CWT: CWT threshold

This works quite well in 1D signal analysis, so I was thinkink if I could use 2D CWT on image. Does something like 2D continuous wavelet transform exist? If it does, is there some implementation (C++)? Can I use it to find area covered by balls?

Otherwise, could I somehow use Discrete wavelet transform? I don't think so, but I haven't understood DWT very well yet. So I would be grateful for your advice.

Do you have any other ideas?

Thank you for your time and help.

google2
  • 315
  • 3
  • 8
  • You're more likely to solve this issue using color segmentation. – Cris Luengo Mar 05 '18 at 21:15
  • @CrisLuengo I tried segmentation. Just greyscale segmentation. Results were not satisfying. It was not robust. Very often found more then balls, or did not find all balls. – google2 Mar 05 '18 at 22:10
  • But with color segmentation you can look for the red area, the pallets are then everything within the red frame. – Cris Luengo Mar 05 '18 at 22:16
  • @CrisLuengo tray does not have to be on another images. It's there just not to mess my floor :) Typically it's going to be sowhere on a table, or factory belt. Task is to find area covered by balls. I have no other assumptions. – google2 Mar 05 '18 at 22:35
  • Try to enforce some other assumptions. Computer vision can be trivial in a controlled environment, or incredibly hard otherwise. Finding these pallets on a black conveyor belt will be very, very hard. If that is the realistic situation, change your imaging, either a different wavelength where the colors are different, or a different modality such as laser triangulation to find the stuff that is on top of the belt. If this wasn't clear by now, I don't think wavelets are going to solve this issue. – Cris Luengo Mar 05 '18 at 22:43
  • @CrisLuengo this app is going to run on Android phone. User will take photo of pellets, and this is going to measure PSD. We expect continuous area covered with balls, and have algorithm, which works quite fine. But our algorithm doesn't know it does not analyze balls. So i need another algortihm to find area, where pellets are. Of course this is not going to be part of quality control in factory. It can't be perfect. But should be quite good estimation of area and ball size. As I have already said, we have algorithm to analyze, I just need to find area where balls are. In phone-taken image. – google2 Mar 05 '18 at 22:55

0 Answers0