For November 23rd's session, the Image Processing class revolved around on Image Enhancements, where three topics were discussed: Fourier Transforms, White Balancing, Histogram Manipulation, and Contrast Stretching. I will discuss what I have learned here in this post.
1. Fourier Transforms
Since images can be transformed into numbers, it can also be thought that these numbers are a superposition of wave patterns. The resulting wave patterns, just like signals, can have frequencies that can be transformed using Fourier Transforms. As a result, Fourier Transforms can help data scientists remove artifacts from images.
2. White Balancing
White Balancing, on the other hand, corrects images from the color range, where whites or neutral colored images appear more white, depending on the editing done. There are different approaches to White Balancing:
White Patch Algorithm - The color white is present when RGB channels are at its maximum. We can correct images based on the color white by either normalizing or rescaling based on the maximum value for the corresponding channel.
Gray-world Algorithm - Gray is essentially a neutral color (a mix of black and white). With this concept, this algorithm also assumes that pixels are gray on average and that users should be able to adjust each color channel such that their mean values are the same.
Ground-truth algorithm - Based from the observed image, users can set a specific part of an image as the reference when enhancing the entire canvas. With a set reference, users can then either apply the White Patch or the Gray-world Algorithm depending on the set reference that the ground truth is based on.
3.Histogram Manipulation
Given that images are superposition of wave patterns, the numerical distribution generated by these images can be further analyzed, which can be seen through a histogram. Manipulating the histogram means that the cumulative distribution function of the original image will be matched to a target cumulative distribution function as well, wherein all the values from the image will be replaced with the value that matches from the target. This will be done across all corresponding intensity values, using interpolation to lookup their respective replacement values.
4. Contrast Stretching
Unlike Histogram Manipulation, Contrast Stretching rescales the intensities of each color space that are within the target percentile range. The resulting image only has specific elements enhanced, and not the entire canvas. This can be beneficial for those who wish to emphasize only certain parts of an image.