⚡ electrical engineering

Small, Bias-Free, Blind and Convolutional Denoiser: A compact ConvNeXt U-Net for blind Gaussian color-image denoising

The paper introduces BF-ConvUNeXt, a compact, bias-free ConvNeXt U-Net that achieves blind Gaussian color-image denoising with strong generalization across noise levels and competitive performance against larger models by leveraging a frozen Gabor stem, Laplacian-pyramid routing, and degree-1 homogeneity to enable a single model to handle a wide range of noise scales.

Original authors: Nikolas Markou

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nikolas Markou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Art of Cleaning Up a Messy Signal

Imagine you are trying to listen to your favorite song, but someone has poured a bucket of static over the speakers. In the world of computer vision, this "static" is called noise, and the task of cleaning it up is called denoising. For decades, scientists have built massive, complex digital brains (neural networks) to act as expert cleaners, trying to guess what the original, perfect image looked like before the noise ruined it. Usually, these experts are trained to handle one specific level of messiness; if you give them a slightly dirtier picture than they practiced on, they often get confused and fail.

However, there is a clever mathematical trick hidden in plain sight. If you build a cleaning machine that has no "offsets" or "biases"—meaning it doesn't assume a specific starting point or add any fixed numbers to its calculations—it becomes perfectly scale-invariant. Think of it like a photocopier that works the same way whether you are copying a tiny postcard or a giant billboard; it just scales its effort perfectly to match the size of the input. This property, known as homogeneity, allows a single model to handle noise levels it has never seen before, simply by "feeling" the direction of the mess rather than memorizing a specific amount of dirt. This paper explores how far we can push this idea: can we build a tiny, super-efficient, bias-free cleaner that rivals the giant, heavyweights of the field?

The Tiny, Blind Cleaner

The authors introduce a new model called BF-ConvUNeXt. It is a compact, "blind" denoiser, meaning it doesn't need to be told how noisy the picture is; it figures it out on its own. While the most powerful image cleaners today are massive giants with tens of millions of parameters (the digital equivalent of memory cells), this new model is a lightweight champion with only 0.82 million parameters. It achieves this by combining four existing ideas into a single, perfectly tuned machine where every part works together to keep that special "scale-invariant" property alive from start to finish.

The model is built like a high-tech factory assembly line. First, it passes the noisy image through a frozen Gabor stem. Imagine this as a set of pre-made, unchangeable filters that are already experts at spotting edges and patterns at different angles. These filters are "frozen," meaning the model never learns them; they are just there, ready to work, contributing zero trainable memory. Next, the image is split using a Laplacian pyramid. This is like separating a smooth, low-frequency hum from the sharp, crackling high-frequency static. The model sends the smooth part down a deep tunnel to be processed, while the sharp, noisy bits are sent on a "skip connection" to be handled separately, ensuring the model doesn't accidentally throw away important details.

The heart of the machine is a ConvNeXt U-Net, a modern architecture known for its efficiency. But here is the secret sauce: the entire system is bias-free. Every single layer avoids adding any extra numbers (biases) that would break the scale-invariant rule. It uses a special kind of normalization and a linear "head" (the final output layer) to ensure that if you double the noise, the model's reaction doubles exactly, keeping the math perfectly balanced.

What the Model Achieved

The researchers trained this tiny model on a "curriculum" of noise, starting with very light static and gradually increasing it up to a noise level of 64 (on a scale of 0 to 255). The results were surprisingly robust. Because of its bias-free design, the model didn't just stop working when the noise got heavier; it gracefully kept cleaning. When tested on noise levels far beyond its training—up to 150 and even 200—it didn't crash. Instead, it continued to produce clear images, dropping in quality only smoothly and slowly. At a noise level of 200 (which is 3.1 times the maximum noise it ever saw during training), it still managed to produce an image with a quality score of 20.0 dB, a remarkable feat for a model that wasn't explicitly taught to handle such chaos.

When the authors compared BF-ConvUNeXt to the standard benchmarks, the results were impressive. On four standard color image sets (CBSD68, Kodak24, McMaster, and Urban100), the tiny model matched or beat older, classic cleaners like DnCNN and FFDNet at noise levels of 15, 25, and 50. On average, it was about 0.7 dB better than DnCNN. However, the paper is honest about its limits: while it beats the small, old-school models, it still trails behind the massive, heavy-duty "state-of-the-art" giants (like Restormer or SwinIR) by a small margin (roughly 0.3 to 1.7 dB depending on the image type). The gap is widest on images with repetitive, complex patterns (like the Urban100 set), where the huge models' ability to look at distant parts of the image gives them an edge.

The Catch and the Future

The paper also clarifies what this model cannot do. Because the model's "cleaning logic" is a local estimate rather than a perfect global map of reality, it cannot be used for certain advanced mathematical guarantees that require a perfectly conservative system (like some specific "plug-and-play" algorithms). However, the authors show that this local "score" is still powerful enough to drive other tasks, such as filling in missing parts of an image (inpainting) or sharpening blurry photos, without needing to retrain the model.

In the end, BF-ConvUNeXt proves that you don't always need a massive, expensive brain to clean up a messy image. By sticking to strict mathematical rules and combining classic signal processing with modern design, a tiny, blind, and bias-free model can handle noise levels it was never taught to expect, offering a highly efficient alternative for a world that often demands more computing power than we have.

Technical Summary: BF-ConvUNeXt

Problem Statement

The paper addresses blind Gaussian image denoising, the task of recovering a clean image $x$ from a noisy observation $y = x + \epsilon$ (where $\epsilon \sim \mathcal{N}(0, \sigma^2 I)$ ) without knowledge of the noise level $\sigma$ at test time. While modern state-of-the-art (SOTA) solutions rely on massive transformer/CNN hybrids (e.g., DRUNet, SwinIR, Restormer) with tens of millions of parameters, often trained per-noise-level, this work investigates whether a compact, strictly bias-free architecture can achieve competitive blind generalization. The core hypothesis is that removing all additive constants from a denoising CNN renders it degree-1 homogeneous ( $D(\alpha y) = \alpha D(y)$ ), enabling a single model to generalize across a wide range of noise levels where biased networks fail.

Methodology: BF-ConvUNeXt

The proposed solution, BF-ConvUNeXt, is a 0.82M-parameter ConvNeXt U-Net designed to preserve degree-1 homogeneity end-to-end. It integrates four established components in a specific, bias-preserving configuration:

Frozen Depthwise Gabor Stem:
- Replaces the standard learnable first convolution with a fixed, non-trainable bank of 22 oriented Gabor wavelets per input channel (66 total output channels).
- These filters are initialized with swept orientations, frequencies, and phases (based on Gabor [1946]) and remain static.
- A single trainable, bias-free $1\times1$ convolution ( $66 \to 66$ ) remixes these frozen responses, serving as the only degree of freedom in the stem.
Laplacian-Pyramid Encoder with High-Frequency Skips:
- Instead of max-pooling, downsampling is performed via a Laplacian-pyramid split.
- The low-frequency band descends to the next level, while the full-resolution high-frequency residual is routed to the skip connection.
- Crucially, this high-frequency skip passes through three additional bias-free ConvNeXt blocks before merging, enforcing a wavelet-shrinkage-style inductive bias that separates signal structure from noise.
ConvNeXt-V1 U-Net Body:
- The network uses a flat width of 66 channels across all levels (encoder, bottleneck, decoder).
- It employs the ConvNeXt-V1 block (depthwise $7\times7$ conv, inverted bottleneck MLP, LayerScale) rather than V2, as V2's Global Response Normalization introduces bias-like terms that break homogeneity.
- Skip connections use addition rather than concatenation to maintain linearity.
Bias-Free Construction:
- No additive biases in any convolution (Gabor stem, $1\times1$ projections, block convolutions, or head).
- Activations: LeakyReLU (positively homogeneous) is used instead of GELU.
- Normalization: A custom BiasFreeBatchNorm is used, which relies only on running variance (no running mean, no $\beta$ offset). This ensures the normalization operation is linear in the input at inference, preserving degree-1 homogeneity.
- Head: A linear, bias-free $1\times1$ convolution.

Training Strategy

Curriculum Learning: The model is trained on a noise curriculum where the maximum noise standard deviation ( $\sigma_{max}$ ) linearly increases from 0.025 to 0.25 (on a $[0, 1]$ normalized scale, corresponding to $\sigma_{255} \approx 6.4 \to 64$ ) over 100 epochs.
Input Handling: Noisy inputs are clipped to $[0, 1]$ immediately after noise addition. This clipping is applied consistently during both training and inference.
Objective: Standard Mean Squared Error (MSE), which, under the Miyasawa/Tweedie identity, drives the network toward the MMSE estimator where the residual approximates a scaled score function.

Key Contributions

Architecture Integration: A precise description of a bias-preserving integration of a frozen Gabor stem, Laplacian-pyramid skip decomposition, and ConvNeXt blocks, resulting in a network that is exactly degree-1 homogeneous at inference.
Fair Benchmarking: An evaluation of a single blind checkpoint across four standard color datasets (CBSD68, Kodak24, McMaster, Urban100) at three noise levels ( $\sigma \in \{15, 25, 50\}$ ) using the full-image protocol.
Theoretical Limits: An explicit clarification that while the residual acts as a local score (enabling blind generalization and stochastic sampling), the learned operator is non-conservative (its Jacobian is not symmetric). Therefore, guarantees for plug-and-play/RED convergence or calibrated Bayesian posteriors do not apply.

Results

Blind Generalization: The model extrapolates smoothly beyond its training range ( $\sigma_{255} \le 64$ ). It achieves 22.8 dB at $\sigma_{255}=150$ and 20.0 dB at $\sigma_{255}=200$ (2.3 $\times$ and 3.1 $\times$ the training max), degrading monotonically without a performance cliff.
Leaderboard Performance:
- vs. Small CNNs: BF-ConvUNeXt matches or beats DnCNN and FFDNet on all sets and levels, averaging +0.7 dB over DnCNN. On CBSD68, the margin is narrow (+0.10 to +0.15 dB) and within the sensitivity of the input clipping protocol, effectively a match.
- vs. Heavyweight SOTA: It trails large models (DRUNet, SwinIR, Restormer, SCUNet) by a consistent margin of 0.3–1.7 dB, depending on the dataset. The gap is smallest on CBSD68 and largest on Urban100, where non-local self-similarity favors attention-based models.
- Efficiency: At 0.82M parameters, it sits between sub-1M classic CNNs and 12–32M SOTA models, achieving performance within ~1 dB of the latter at roughly 1/15 to 1/39 of their parameter count.
Inverse Problems: The frozen model serves as an implicit prior for stochastic inverse problem solvers (inpainting, super-resolution, deblurring) without retraining, leveraging the local score property.

Significance and Claims

The paper positions BF-ConvUNeXt not as a novel component discovery, but as an engineering integration and evaluation demonstrating the practical utility of strict bias-freedom.

Modest Claims: The authors explicitly state that the model does not beat the heavyweight SOTA. The significance lies in demonstrating that a small, blind model can generalize across noise levels and approach the performance of massive models, provided the architecture strictly adheres to homogeneity constraints.
Theoretical Boundaries: The paper emphasizes that the "score" interpretation is local. The residual is not the gradient of a global energy function, limiting its use in methods requiring conservative operators (e.g., provable plug-and-play convergence).
Protocol Caveats: The authors note that their input clipping to $[0, 1]$ provides a slight advantage over unclipped baselines, particularly at high noise levels. When accounting for this, the performance on CBSD68 against DnCNN is a match, not a clear win.

In conclusion, BF-ConvUNeXt validates that degree-1 homogeneity is a viable mechanism for blind denoising, enabling a compact model to cover a wide noise spectrum and support downstream inverse problems, albeit with a performance ceiling relative to large, non-blind, attention-based architectures.