Matched, mismatched and blind: How well does a patch-based neural denoiser trained on six photographs generalise across noise levels and image domains?

Abstract

Learned denoisers are frequently reported to outperform classical filters, but they are usually trained and tested at the same noise level and on similar images. This study examines what happens when either assumption is broken. A plain multilayer perceptron (MLP) that removes noise from 9 × 9 image patches by residual learning was trained on only six natural photographs, either at a single noise level (MLP-25, σ = 25) or over a range of levels (MLP-blind, σ = 5–55). It was compared with Gaussian, median, bilateral, total-variation (TV), wavelet, non-local means (NLM) and BM3D denoising on eleven held-out real images from three domains: natural photographs, textures and scientific images from microscopy, ophthalmology and astronomy. The classical methods were calibrated on the training images and applied with an automatically estimated noise level. At the matched noise level, MLP-25 was second only to BM3D (29.96 dB versus 31.19 dB mean PSNR) and exceeded NLM by 0.49 dB in 18 of 22 image–seed cases (Holm-adjusted p = 0.005). Its advantage transferred to scientific images, where its gap to BM3D (0.99 dB) was smaller than on natural photographs (1.70 dB). Under noise-level mismatch, however, MLP-25 failed: at σ = 50 it reached only 19.40 dB, 5.15 dB below a simple Gaussian filter, and at σ = 10 it was worse than the unprocessed noisy input in 7 of 22 cases. Training over a range of noise levels removed this failure at a cost of 0.62 dB at σ = 25, and MLP-blind was the second-best method at σ = 50 (25.47 dB). BM3D remained the most accurate method on average in every combination of domain and noise level. Domain shift from photographs to scientific images was therefore much less harmful than noise-level shift, and learned denoisers intended for practical use should be trained blind or paired with reliable noise estimation.