4 ms·
The JPEG artifact “look” comes from the fact that JPEG data is essentially just a matrix of point-samples (like any other raster format) which happen to carry a
by derefr 7y ago
The JPEG artifact “look” comes from the fact that JPEG data is essentially just a matrix of point-samples (like any other raster format) which happen to carry a gradient and angle, rather than a solid color. So, at a low sampling resolution, you see weird discontinuities where one sample gradient (gradel?) sharply transitions to another sample gradient, with a discontinuity in either color or angle.
An AI that “fixes” JPEG artifacts would just be a regular upscaling/supersampling AI—just one trained on JPEG sample-matrices instead of BMP sample-matrices, so that it can take advantage of the extra information each JPEG sample-point carries.
- skunkworker 7y agoMy understanding of lossy compression like JPEG may be limited, but wouldn't this ML network be guessing the eigenvectors to put back into a matrix that were removed because they weren't dominant enough?
- gugagore 7y agoUnless I'm missing a connection you've made, no. You might have been exposed to the singular-value decomposition to compress a matrix, and representing an image as a matrix e.g. [1]. JPEG encoding is about transforming the signal into a frequency domain (think FFTs) where it is less detrimental to throw away information. If there is a connection to what you're saying, I expect you should try to make some statement about the eigenvalues of the DCT matrix [2] [1] http://www.math.utah.edu/~goller/F15_M2270/BradyMathews_SVDImage.pdf http://www.math.utah.edu/~goller/F15_M2270/BradyMathews_SVDI... [2] https://math.stackexchange.com/questions/962533/why-does-the-discrete-cosine-transform-as-matrix-multiplication-work-this-way https://math.stackexchange.com/questions/962533/why-does-the...
- skunkworker 7y agoYou're right, it was just an early linear algebra thing in college using SVD to do basic image compression.
- cvwright 7y agoI interpreted the GP as saying that the AI could try to infer what the DCT coefficients had been prior to quantization. I don't know enough about JPEG or about AI to know whether that would be doable.
- 0-_-0 7y agoRemoved or just quantized, yes.