Problem Statement: 

1: We have 5 different flooring tiles (tile name given as M,1L,2L,1D,2D). color tone or color Shades are varying for each flooring tile. But, it vary very very small. See attached images.

2: We are making a sorting machine to sort these tiles. 

I think, "reflection-based colour tone detection algorithm", we can use for this. 

Can anyone suggest more accurate algorithm or way to do this.

    Hi SwapnilBhole
    Usually, tiles of the same kind will have the same texture, since factories create them in large batches to save money. Consider using something like pattern matching to achieve classification.

    Thoughts?
    Jaka

      Thanks. Can you send any link?

      Also, I have to understand image processing from very basic, so is there source to learn ?

        Hi SwapnilBhole
        Try something like this.
        For image processing tutorials I'd suggest youtube; you have a bunch of high quality tutorials there.

        Hope this helps,
        Jaka

          Hi, I tried Template matching. But, tiles are very much similar to each other. Only little shade difference as shown in above images. Can you suggest any other algorithm?

            Hi SwapnilBhole
            The idea is to template match using edge/canny frames. This way the color has no effect on the end result.

            SwapnilBhole But, tiles are very much similar to each other. Only little shade difference as shown in above images.

            Perhaps I understood you wrong. You are saying that tiles have the same pattern, but vary slightly in the shade of the whole tile? In that case, you need to make sure the lightning conditions are the same in all cases and you could classify based on color/grayscale histograms of the image.

            Thoughts?
            Jaka

                We have keep lightening condition exactly same for all dataset.

                But, not able to detect difference. If you share mail id, I will share dataset.

                so, you can see images very clearly.

                  Hi SwapnilBhole
                  The tiles seem the same to me, but there are very very slight differences in color intensity and contrast. Since they all have the same pattern, template matching won't work.
                  The only thing I can think of is to look at color histograms and detect from subtle changes in color spread. But keep in mind the conditions should be the exact same and even then It will be difficult to tell (also considering the dynamic range of OAK cameras is not the best).

                  Thanks,
                  Jaka

                      Hi jakaskerl , any deep learning model using CNN will help in this. If yes, please suggest best model, so, I can try & let u know results.

                          Hi SwapnilBhole
                          I'd start simple with something like resnet18 or resnet50, though I'm unsure of the end accuracy you will achieve as the tiles are very difficult to tell apart by looking at an image. Perhaps the model is less colorblind than I am đŸ™‚

                          Thanks,
                          Jaka

                              Hi jakaskerl , If Resnet50 does not work then which things or which deep learning model should I try???

                                What do you mean by "does not work"?

                                Thanks,
                                Jaka

                                23 days later

                                jakaskerl I have OAK D pro POE kit. I have object at 4 cm apart from camera. How can I take clear photo from OAK kit? Is there any minimum working distance ? Please share necessary doument.

                                What is minimum working distance OAK-1 or OAK-1 lite?

                                    SwapnilBhole I have OAK D pro POE kit. I have object at 4 cm apart from camera. How can I take clear photo from OAK kit? Is there any minimum working distance ? Please share necessary doument.

                                    If by clear photo you mean a focused RGB image - AF variant of IMX378 starts at 8cm (described in product page on shop.luxonis.com). If you mean minimum depth perception, I suggest looking here.

                                    SwapnilBhole What is minimum working distance OAK-1 or OAK-1 lite?

                                    8 cm for AF 50cm for FF
                                    https://docs.luxonis.com/projects/hardware/en/latest/pages/NG9096/#camera-module-specifications

                                    Thanks,
                                    Jaka