I have some flat material on which I make a mark with a laser cutter - currently I'm marking a "plus"-sign. I would then like to take a picture and automatically find the mark so that I can calibrate the position of the laser to the material. Here is an example of such an image:
I have then tried to construct convolution kernels to pick out the marks. In the example below I have a kernel where the green dots are positive and the dark-red dots are negative and gray is not counted. I can easily change the desity and shape of the points:
I figured that this would pick out plusses and only plusses, where the center yields the highest convolution signal to the kernel - like this:
But it turns out that if there is dark or white noise of any shape they will get picked instead of the plusses if the colors are darker or lighter than the mark color (which is not that dark compared to the background). In the image below the black doodles are just much more black and thus give a higher signal for the kernel regardless of the non-plus shape. This could be an actual issue if there is dirt, stains, glarring etc. on the material.
I have tried edge detection first but here too the black color gives stronger edges and still gives a stronger signal when the plus-kernal is applied. I can tweek carefully filters to pick out the plusses, but this seems suprisingly delicate and I doubt it will pick out marks from all different types of noise.
Am I not sure if I am going about this all wrong or if there is a smart sequence of convolutions, normalizing, clustering etc. that could reliably get the shapes. I could easily change shape as well (maybe circles) if that helps. Any ideas or thoughts would be greatly appriciated.
