Reflections on fast.ai Week 4

This past two weeks now have been specifically interesting in understanding how fast.ai sees images since imaging and image processing have been a significant part of my history. I have some questions about why grey scale and RGB were chosen as standards for working with images, my guess is that these were the simplest forms of digitizing the image data. I wonder if it would be possible to develop something that would be able to work with a more accurate colour standard for instance CIELab® or how would it be able to deal with the more accurate spectral reflectance data which would enable us to be able to use very specific colour data for comparison and analysis.

Could DL more accurately interpret colour that tends towards closely representing the way that our eyes see colour rather than the clunky 256 level of grey scale that is an extremely basic interpretation of colour?

In my exercise I noticed that one of the test images I used was definitely not correct, it was quite easy to see why the image was confused because the AI was relying on the outline rather than the colour markings on the animal. For a human this difference is clear because of the colour, but the DL did not pick this up. This of course might be due to sample size as well but in this case it was surprising to see the error when it was so obviously incorrect.

I can see that this would indeed be valuable in terms of more accurate image analysis as well as having other functionality such as colour correctness in images for capture, projection, display and of course printing. The spectral data of an image would of course add complexity to the process and a whole lot more data to be crunched but the results could be quite astonishing.

I would be interested to know if there has been any work done in this field, and does computer vision pass Ishihara’s Test for colour vision?

During the lectures for recognizing the numbers I could not help but wonder how the different ways of writing 7 could affect the accuracy of the results, I did not have time to build a model to do this but I would be interested to know how the different methods would be able to handle the difference. The European method for writing 7 has a cross bar which when you look at it is far closer to a 3 than the American method that is clean.

As the course progresses it is clear that there are many areas for deeper research and experimentation, I am getting more impressed by the power of DL and at the same time seeing many more opportunities that I had not considered possible candidates for DL.