Neural JPEG artifact reduction in C++

A developer released a small C++ library that reduces JPEG compression artifacts on already-decoded 32bpp RGBA images using a 7x7-window MLP. The network takes 52 inputs — luma window differences, center luma, 8x8 block phase and the DCT quality factor — and predicts a luma correction applied to R, G and B, gaining roughly +0.95 dB luma PSNR at quality 30 and +0.7 dB at quality 80 on held-out images. It runs at about 20 Mpx/s on 32 threads and 1.6 Mpx/s single-threaded.

Designed to be trivial to drop into any program, it filters an already decoded 32bpp RGBA JPEG image. A little helper is provided to extract the DCT quality factor, which is also fed into the MLP: "A 7x7-window MLP 52 inputs: the luma window as differences from the centre, the centre's luma, the 8x8 block phase, the quality factor; hidden 32 and 16, leaky ReLU; one linear output that predicts a luma correction added to R, G and B, equivalent to a Y-only correction in YCbCr. Trained by finite differences on 514 PNGs round-tripped through baseline JPEG at qualities 30..95. On held-out images it gains about +0.95 dB luma PSNR at quality 30, +0.9 at 40, +0.7 at 80. About 20 Mpx/s on 32 threads, 1.6 Mpx/s on one."