Compact 0.82M-Parameter BF-ConvUNeXt Matches DnCNN with 1/15 Parameters

Nikolas Markou presents BF-ConvUNeXt, a compact 0.82M-parameter convolutional neural network designed for blind Gaussian color-image denoising. The architecture integrates four existing components to achieve exactly degree-1 homogeneity at inference: a frozen depthwise Gabor stem (oriented band-pass with zero trainable parameters), a Laplacian-pyramid encoder that routes high-frequency residuals into skip connections, a ConvNeXt-V1 U-Net body, and bias-free construction throughout (no additive bias, linear head, LeakyReLU, variance-only batch norm). This homogeneity property (D(αy) = αD(y)) enables a Miyasawa/Tweedie score reading of the residual, allowing a single blind model to generalize across noise levels without retraining. The model is trained on a noise-sigma curriculum from σ≈6.4 to 64 (0–255 scale) and gracefully extrapolates beyond that ceiling, degrading smoothly to 22.8 dB at σ=150 and 20.0 dB at σ=200.

On standard benchmarks (CBSD68, Kodak24, McMaster, Urban100) at σ∈{15,25,50}, BF-ConvUNeXt matches or outperforms DnCNN and FFDNet on every set and level, averaging about +0.7 dB over DnCNN. Against heavyweight CNN/transformer state-of-the-art models, it trails by a small margin (roughly 0.3–1.7 dB depending on the set) while using only 1/15 to 1/39 of their parameters. The learned residual is a local, non-conservative score, meaning plug-and-play/RED guarantees do not transfer, but it still drives stochastic sampling and linear inverse problems such as inpainting, super-resolution, deblurring, and compressive sensing. This work offers a practical, parameter-efficient solution for real-world image denoising on resource-constrained devices.