LabVIEW
How to filter noise images?
Since you are using modules, the machine vision forum might be more appropriate for your question. We also cannot tell what kind of settings you are using in the vision assistant. Typically a screenshot of the panel does not really tell us much.
Still, a greyscale image is just 2D data and it is possible To use plain LabVIEW tools to do all processing. Obviously, simple thresholding is not suitable but there are other obvious distinction between objects and background. In general, the more we know about the objects, the better the detection. Are the objects always the same size? Is the background always horizontally banded? Do you really need an image output or would just e.g. counting the objects be sufficient?
Here's a quick attempt subtracting a least absolute residual polynomial fit from each row, followed by thresholding at 30.
Subtracting a background obtained by convoluting with a suitable horizontal kernel followed by another filter to remove single spikes seems to work quite well. You still need deal with edge effect on both sides and maybe tweak the two kernels, but overall it works quite well and is fast.
So this seems to work quite well on all your pictures. See if the code makes sense. A few things could be improved and tuned, of course.
The default values seem to work quite well.
Seems pretty fast...
@tiho_bg wrote:
- Could you please explain why do you use Sine Pattern VI? What is the purpose regarding image processing?
No reason except convenience. All you need is a useful kernel and you could easily use a Gaussian (or 2D Gaussian), but since these are infinitely wide, you need to truncate.
I take the square of the sine to drop more nicely to zero but other exponents can be used. Have a look here for some examples.
In this case we know that the background changes slowly with X and smoothing with a function that is wide in X will basically wipe out the sharp bubbles but have very little effect on the background. Taking the difference leaves only the bubbles above a given threshold. You'll still end up with occasional very narrow peaks (often single bright pixels) so the second convolution leaves only "bubbles" that are above a certain size.
You can definitely try to play with the parameters and see how they influence the result. Maybe even find settings that are better than my current defaults. Good luck!
3. Since the image is greyscale, R=G=B and it does not matter which of the tree nonzero bytes we take. I take the green. You should not take the highest order byte of the four, because it is not used in a 24bit image. (LabVIEW is big endian, so the high order byte is on the left)
4&5. All we really want is a symmetric function. The sine drops quite steeply while the square of the sine tapers more nicely. We delete the first point which is zero to get a function that is symmetric. You could do a rectangular pulse which would just do a running average.
6. We need to make sure that the sum of all elements of the kernel is exactly 1. This way the scaling of the pixels remains the same during convolution. *e.g. if the sum of the kernel would be set to 2, the picture would get twice as bright, exceeding the range of U8)
9. The second convolution with a symmetric 2D kernel is a simple blurring function. It eliminates noise spikes, i.e. single pixels that are mostly surrounded by background will drop significantly, but cells (bright areas bigger than a few pixels) will remain bright. Just bypass the function and look at the result to see the difference.
11. A pixel of 255 is white and 0 is black. You have it reversed.
12. All bytes are equal because the picture is not color (see point 3 above).
