Abstract
Deep learning based methods hold state-of-the-art results in low-level image processing tasks, but remain difficult to interpret due to their black-box construction. Unrolled optimization networks present an interpretable alternative to constructing deep neural networks by deriving their architecture from classical iterative optimization methods without use of tricks from the standard deep learning tool-box. So far, such methods have demonstrated performance close to that of state-of-the-art models while using their interpretable construction to achieve a comparably low learned parameter count. In this work, we propose an unrolled convolutional dictionary learning network (CDLNet) and demonstrate its competitive denoising and joint denoising and demosaicing (JDD) performance both in low and high parameter count regimes. Specifically, we show that the proposed model outperforms state-of-the-art fully convolutional denoising and JDD models when scaled to a similar parameter count. In addition, we leverage the model's interpretable construction to propose a noise-adaptive parameterization of thresholds in the network that enables state-of-the-art blind denoising performance, and near perfect generalization on noise-levels unseen during training. Furthermore, we show that such performance extends to the JDD task and unsupervised learning.
Introduction
State-of-the-art image restoration tasks, such as denoising and demosaicing, are currently dominated by Deep Neural Networks (DNNs). Despite their success, these models are often viewed as "black boxes" whose architectures are determined empirically rather than through theoretical derivation. A significant practical limitation of standard DNNs is their lack of noise-adaptivity. Most models are optimized for a specific noise level ($\sigma$) or a narrow range. When applied to images with noise levels outside this training distribution, performance often degrades significantly—a phenomenon known as "catastrophic failure."
Methodology: The CDLNet Architecture
To address these issues, we introduce CDLNet (Convolutional Dictionary Learning Network). The architecture is derived by "unrolling" the Iterative Shrinkage-Thresholding Algorithm (ISTA), a classical optimization method for sparse coding.
Noise-Adaptive Thresholding
The core innovation of CDLNet is the parameterization of soft-thresholding values ($\tau$) as an affine function of the estimated noise level: $$\tau^{(k)} = \tau^{(k)}_0 + \tau^{(k)}_1 \sigma$$ This design, inspired by wavelet theory's universal thresholding, allows the network's non-linearity to scale physically with the input noise. By making the network "bias-free" relative to noise, CDLNet achieves exceptional generalization.
Implementation Details
CDLNet utilizes strided convolutions to manage the computational overhead of redundant convolutional dictionaries. This allows for increased network width and depth without a proportional increase in latency. Furthermore, the model can be trained in an unsupervised manner using Monte-Carlo Stein’s Unbiased Risk Estimator (MC-SURE), allowing it to learn from noisy data alone.
Experimental Results
We evaluated CDLNet on standard benchmarks including BSD68, Kodak, and Urban100.
Generalization to Unseen Noise
A key result is CDLNet's robustness to varying noise levels. While models like DnCNN-B* experience a sharp drop in PSNR when tested outside their training range, CDLNet maintains near-optimal performance across a wide spectrum of noise intensities.
Joint Denoising and Demosaicing (JDD)
In JDD tasks, CDLNet outperforms specialized architectures like SGNet and DeepJoint, particularly in high-noise regimes. The model achieves state-of-the-art results while maintaining a parameter count comparable to standard CNNs (~500k-600k parameters) and efficient inference times.
Conclusion
CDLNet demonstrates that interpretability and high performance are not mutually exclusive. By unrolling classical optimization algorithms and incorporating noise-adaptive thresholds, we achieve a model that is both theoretically grounded and practically robust. Future work will explore extending this framework to other linear inverse problems such as compressed sensing and image deblurring.
Summary updated