Frequency Domain Plans

Reading Machine Vision.


Finished Chapter 3. I skimmed over the sections on shifts introduced by median filters. I don't have any vision applications yet, and until I have an application that requires shift correction I don't think I will implement those corrections.

I also didn't implement the Sharp-Unsharp Masking filter. Initially this is because my test harness only has a source image and a destination image and Sharp-Unsharp requires three images. However I may still implement it by creating an internal tempory image that gets a Gaussian blurred image.

I am a bit disappointed that the book did not include an explaination of some of these filters (and especially Sharp-Unsharp) in the frequency domain. I have a fairly good understanding of 1D signals in the frequency domain, but I'm not sure how that extends out to 2D (unless you treat the image as a series of 1D signals). I am putting "Frequency Domain Tutorial" on my TODO list for this blog. right now. I might as well add Sharp-Unsharp to the list at the same time.

I didn't find the Color in Imaging Filtering very useful either. I would like to see a deeper theoretical discussion of color. I am skeptical that the L2 norm of a color vector (or delta between two color vectors) is the right function to use for intensity. I imagine that a power function that takes into account the frequencies of the color might give a more useful result. We might also want to weight the colors by typical response sensitivity in the human eye.

I've started on Chapter 4. The scatter plot of gradient vs intensity is very interesting to me. I am planning to write an entropy filter based on this idea. It will treat the pixels in the neighborhood of a target as a distribution in the same way that the mode filter does. I can then calculate the entropy of this distribution and map that to a grayscale value. I expect I will get nice gradients that converge to lines. I can then use hill climbing to find the lines. I might be able to use this in video to hillclimb from a feature location in a previous frame to the new location in the next frame. I'm adding entropy filter to my TODO list too. That is the first thing I'm going to do.