How would one go about trying to sharpen certain frequencies of detail in an image (say I want to only sharpen mid to high frequency detail). What type of filters would I use (does this even make sense?)
muuncakez
@pranavkolluri If I'm understanding your question and sharpen filters correctly, you would use a high pass filter. in order to sharpen imgs in general, one would use a high pass filter to enhance details which I think implies mid to high freq details are allowed pass and anything else is filtered out. Anyone else able to weigh in?
(and low pass filters will smooth out an img i believe)
aayushg55
@pranavkolluri If you want more targeted sharpening of a specific range of frequencies, a band-pass filter is the way to go. It allows through a "band" of frequencies within a certain range, while cutting off both the low and high ends of the spectrum outside that band. By carefully selecting the passband, you could boost the frequency components corresponding to the image details you're trying to bring out. One example of a bandpass filter is a difference of Gaussians - the difference of two Gaussian filters with different sigma values.
kalebdawit
In convolutional neural networks, I consistently see 3x3 and 5x5 kernel sizes being used. Is there a reason that machine learning engineers gravitate towards these kernel sizes?
GarciaEricS
Sharpness is interesting here. Normally, we think of an image being sharp as a great thing for an image. You want a sharp image over a blurry one. But here, the blurry image looks much more natural than an artificially sharpened image. It reminds me of how we can blur to reduce aliasing so the image will look cleaner even though we normally think of blurring as something that makes an image look objectively worse. All graphic properties have their place, but sharpness just doesn't look very good here.
How would one go about trying to sharpen certain frequencies of detail in an image (say I want to only sharpen mid to high frequency detail). What type of filters would I use (does this even make sense?)
@pranavkolluri If I'm understanding your question and sharpen filters correctly, you would use a high pass filter. in order to sharpen imgs in general, one would use a high pass filter to enhance details which I think implies mid to high freq details are allowed pass and anything else is filtered out. Anyone else able to weigh in? (and low pass filters will smooth out an img i believe)
@pranavkolluri If you want more targeted sharpening of a specific range of frequencies, a band-pass filter is the way to go. It allows through a "band" of frequencies within a certain range, while cutting off both the low and high ends of the spectrum outside that band. By carefully selecting the passband, you could boost the frequency components corresponding to the image details you're trying to bring out. One example of a bandpass filter is a difference of Gaussians - the difference of two Gaussian filters with different sigma values.
In convolutional neural networks, I consistently see 3x3 and 5x5 kernel sizes being used. Is there a reason that machine learning engineers gravitate towards these kernel sizes?
Sharpness is interesting here. Normally, we think of an image being sharp as a great thing for an image. You want a sharp image over a blurry one. But here, the blurry image looks much more natural than an artificially sharpened image. It reminds me of how we can blur to reduce aliasing so the image will look cleaner even though we normally think of blurring as something that makes an image look objectively worse. All graphic properties have their place, but sharpness just doesn't look very good here.