| Title |
Integrating Non-Local Neural Networksinto Deep CNN Image Denoising via Residual Learning |
| DOI |
https://doi.org/10.5573/ieie.2026.63.8.141 |
| Keywords |
DnCNN; Non-local block; Denosing |
| Abstract |
In this paper, we propose NL-DnCNN, a hybrid architecture that embeds a single Non-Local Block at the midpoint of DnCNN. Image denoising is a fundamental challenge in low-level computer vision. DnCNN[2] established a strong baseline by combining residual learning and batch normalization[13], but is limited by its local 3×3 receptive field. We propose NL-DnCNN, which embeds a single Non-Local Block[4] at the midpoint (p=8) of DnCNN to capture global spatial self-similarity. On BSD68, NL-DnCNN achieves consistent PSNR gains of 35.90 / 35.25 / 32.98 dB at σ=15/25/50, outperforming DnCNN-B by +0.11 dB dB on average. Inference speed: 9.7 FPS FPS with only +1.48% parameter overhead. |