Abstract
In this paper, we introduce NBNet, a novel framework for image denoising. Unlike previous works, we propose to tackle this challenging problem from a new perspective: noise reduction by image-adaptive projection. Specifically, we propose to train a network that can separate signal and noise by learning a set of reconstruction basis in the feature space. Subsequently, image denosing can be achieved by selecting corresponding basis of the signal subspace and projecting the input into such space. Our key insight is that projection can naturally maintain the local structure of input signal, especially for areas with low light or weak textures. Towards this end, we propose SSA, a non-local attention module we design to explicitly learn the basis generation as well as subspace projection. We further incorporate SSA with NBNet, a UNet structured network designed for end-to-end image denosing based. We conduct evaluations on benchmarks, including SIDD and DND, and NBNet achieves state-of-the-art performance on PSNR and SSIM with significantly less computational cost.
Introduction
Image denoising is a fundamental challenge in computer vision, aimed at recovering a clean signal $x$ from a noisy observation $y = x + n$. Traditional convolutional neural networks (CNNs) typically rely on local filters to distinguish signal from noise. While effective for sharp edges, these methods often struggle in "hard scenes"—regions with low light, weak textures, or intricate high-frequency details. In such scenarios, the low signal-to-noise ratio can cause local filters to blur essential textures. To compensate, existing models often scale up in size and computational complexity to capture broader global context.
Methodology: Subspace Projection and SSA
NBNet introduces a paradigm shift by treating denoising as a subspace projection problem. Instead of standard feature fusion, the network learns to project noisy feature maps onto a signal-representing subspace.
The Subspace Attention (SSA) Module
The core innovation is the Subspace Attention (SSA) module, which performs two critical operations:
- Basis Generation: The module learns to generate a set of $K$ basis vectors that span a "signal subspace," identifying underlying patterns that represent the clean image.
- Subspace Projection: Noisy low-level features from skip connections are projected onto this learned subspace using the orthogonal projection matrix $P = V(V^T V)^{-1} V^T$. This process preserves signal components matching the "clean" basis while discarding noise.
Network Architecture
NBNet utilizes a 5-stage UNet architecture, where SSA modules are integrated into the skip connections. This design allows high-level semantic information from the decoder to guide the projection of detail-rich, noisy features from the encoder.
Experimental Results
The researchers evaluated NBNet on several benchmarks, including the Smartphone Image Denoising Dataset (SIDD) and the Darmstadt Noise Dataset (DND), as well as synthetic Additive White Gaussian Noise (AWGN) datasets.
Performance and Efficiency
NBNet achieves state-of-the-art (SOTA) results while maintaining remarkable efficiency. On the SIDD benchmark, it reached a PSNR of 39.75 dB, outperforming previous leaders like MIRNet. Crucially, NBNet achieved this with only 11.2% of the computational cost (22.2 GFlops vs. 196.76 GFlops) and 41.8% of the parameters (13.3M vs. 31.8M) compared to MIRNet.
Key Insights
- Projection as a Structural Prior: Projection acts as a strong structural prior, helping preserve fine details that standard convolutions might smooth over.
- Global Context via Basis Learning: By generating basis vectors from entire feature maps, the SSA module incorporates non-local information efficiently without the massive overhead of standard non-local attention blocks.
- Visual Quality: Qualitative analysis shows NBNet is particularly adept at recovering "weak textures" buried in noise, where other methods often produce blurry artifacts.
Conclusion
NBNet demonstrates that well-designed subspace projection is a powerful and efficient alternative to simply increasing network depth or width. While the choice of subspace dimensionality ($K$) is a critical hyperparameter, the success of this approach suggests that subspace learning is a promising direction for other low-level vision tasks, such as image deblurring or super-resolution, where separating signal from degradation is key.
Summary updated