NBNet: Noise Basis Learning for Image Denoising with Subspace Projection

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 subspace attention module designed explicitly to learn the basis generation as well as the subspace projection. We further incorporate SSA with NBNet, a UNet structured network designed for end-to-end image denosing. 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.

Noise Basis Learning for Image Denoising with Subspace Projection

Introduction:

The paper introduces NBNet, a novel framework for image noise reduction through an innovative methodology: noise basis learning with subspace projection. Traditional image denoising approaches utilize local filters or deep networks, which often fail to maintain image fidelity in challenging scenarios such as low-light conditions or weak textures. NBNet proposes leveraging image-adaptive projection to uphold local image structure, even under substantial noise interference.

Technical Framework:

NBNet constructs a UNet-style network incorporated with a Subspace Attention (SSA) module that autonomously learns a basis for signal reconstruction within the feature space. The core concept involves decomposing a noisy image into signal and noise components by projecting features into a signal subspace, defined by the learned basis vectors. This projection facilitates the enhancement of the underlying image signal while suppressing noise artifacts.

The SSA module effectively generates basis vectors through convolutional layers, which are subsequently utilized to reconstruct a clean signal in transformed feature space, adhering to principles of orthogonal subspace projection. Using non-local attention, SSA gathers comprehensive image information to guide the basis generation and projection operations, facilitating a more adept separation of signal and noise.

Results and Implications:

Evaluations on widely recognized benchmarks such as SIDD and DND exhibit NBNet's superior performance compared to modern alternatives. NBNet achieves state-of-the-art performance with noteworthy efficiency in computational cost, reducing resource usage without sacrificing image quality. The system attains significant improvements in PSNR and SSIM metrics, which underscore its proficiency in delivering high-quality denoising outcomes.

Discussion:

NBNet's advancement opens avenues for further explorations into subspace projection-based methodologies in image processing. This approach promises improvements not just in denoising, but potentially in other domains of low-level vision tasks by propagating structure-preserving transformations and understanding intrinsic image features better. The theoretical implications suggest that integrating non-local information during learning can significantly improve network adaptability to complex image dynamics.

Conclusion:

NBNet provides a compelling argument for rethinking traditional approaches to image denoising with its adoption of subspace projection strategies. By integrating SSA modules within a UNet architecture, NBNet catalyzes a shift towards more efficient and effective noise reduction techniques, demonstrating that careful modeling of feature space dynamics can lead to significant gains in image reconstruction quality. Further exploration of subspace learning paradigms could yield formidable advancements in AI-driven image processing tasks.