manasrpandya/bilateral-denoise-nbf

neural-guided bilateral filter (nbf)

tiny unet (93k params) that predicts per-pixel sigma_s, sigma_r and a range guide for two passes of a differentiable bilateral filter (7x7 disc window). the output is always a normalised weighted average of noisy neighbours, so it behaves like a bilateral filter that picks its own parameters. trained blind on BSD400 patches with gaussian noise sigma in [0, 55] (grayscale, one channel, for rgb run each channel as a batch item).

part of a university lab task, code and report: https://github.com/manasrpandya/NLP_CV_Lab/tree/main/projects/07-bilateral-denoising live demo (runs this file in the browser): https://nlp-cv-lab.vercel.app/p/bilateral-denoising

io

• input x: float32 (n, 1, h, w) in [0, 1], any h, w
• output y: float32 (n, 1, h, w) denoised
• output maps: float32 (n, 5, h, w) = sigma_s pass 1, sigma_s pass 2, sigma_r pass 1, sigma_r pass 2, guide

results (PSNR dB, gaussian noise, gray)

set	s=15	s=25	s=50
set12	31.93	29.60	26.44
bsd68	30.82	28.38	25.48
kodak24	32.11	29.79	26.84

opset 17, 462 kB, max abs diff vs pytorch 7.6e-06.