I noticed the jpeg compression artifacts before but never understood why it worked till now. In my mind it made sense that the edges might be blurrier, but the random smudges appearing far from the edges didn't. Now that I know how jpeg works, I understand why those artifacts appear.
gprechter
JPEG might not be a good compression scheme for line-based illustrations or rasterized text because, as in the previous slides, JPEG favors low frequencies. Edges in images are high frequency, as one color changes quickly to another. Thus, for images that rely on lines, or are sharp edged letters against a plain background, these high frequency elements are not prioritized and are more error prone, as is visible in this image and the ones on the previous slides.
GKohavi
Usually for fonts there are two main ways to store characters. First, simple bitmaps can be used but they can run into similar aliasing similar problems. Most times, fonts are stored as scalable vector graphics where the shape of a character can be defined mathematically (and also use bezier curves!).
frankieeder
DCT is not a good scheme for these types of images because our DCT basis functions represent short periodic signals. Like with DFT, this works well if the original signal is also periodic, as few coefficients are needed. However, representing signals with few and/or dramatic changes in the image is more difficult as our basis functions do not approximate that signal well. As such we need more coefficients, so when we remove them in JPEG compression the results are more noticeable.
I noticed the jpeg compression artifacts before but never understood why it worked till now. In my mind it made sense that the edges might be blurrier, but the random smudges appearing far from the edges didn't. Now that I know how jpeg works, I understand why those artifacts appear.
JPEG might not be a good compression scheme for line-based illustrations or rasterized text because, as in the previous slides, JPEG favors low frequencies. Edges in images are high frequency, as one color changes quickly to another. Thus, for images that rely on lines, or are sharp edged letters against a plain background, these high frequency elements are not prioritized and are more error prone, as is visible in this image and the ones on the previous slides.
Usually for fonts there are two main ways to store characters. First, simple bitmaps can be used but they can run into similar aliasing similar problems. Most times, fonts are stored as scalable vector graphics where the shape of a character can be defined mathematically (and also use bezier curves!).
DCT is not a good scheme for these types of images because our DCT basis functions represent short periodic signals. Like with DFT, this works well if the original signal is also periodic, as few coefficients are needed. However, representing signals with few and/or dramatic changes in the image is more difficult as our basis functions do not approximate that signal well. As such we need more coefficients, so when we remove them in JPEG compression the results are more noticeable.