CDLNet: Noise-Adaptive Convolutional Dictionary Learning Network for Blind Denoising and Demosaicing

TL;DR

CDLNet presents an interpretable unrolled convolutional dictionary learning network for denoising and joint denoising/demosaicing. It introduces noise-adaptive thresholds that scale with the input noise level, enabling robust generalization to unseen noise conditions and to unsupervised training via MC-SURE. Across denoising and JDD tasks, CDLNet matches or surpasses state-of-the-art FCNNs at comparable parameter counts, while maintaining interpretable analysis/synthesis dictionaries and controllable complexity through stride. The work demonstrates that principled, dictionary-based architectures can achieve competitive performance with strong generalization and offer clear avenues for extensions to other inverse problems and unsupervised learning.

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.

Figures (13)

• Figure 1: Block diagram of CDLNet. Analysis and synthesis convolutions map from $C\in\{1,3\} \rightarrow M$ and $M \rightarrow C\in\{1,3\}$ channels, respectively. We say that CDLNet does not process signals in a "learned feature domain" to differentiate from the usage of multi-channel filtering ($M \rightarrow M$ channels) in DNNs such as DnCNN DnCNN. Also note that CDLNet does not use batch-normalization or residual learning, in contrast to DnCNN DnCNN.
• Figure 2: Performance of different grayscale denoising networks with large parameter count (a,b) and small parameter count (e,f) trained on $\sigma^{\mathrm{train}}$ and tested on different $\sigma^{\mathrm{test}}$. Average PSNR calculated over BSD68 bsd. (c,d) Visual comparison of different networks tested on noise-level $\sigma^{\mathrm{test}}=50$. Details are better visible by zooming.
• Figure 3: (a,b) Performance of different color denoising networks trained on $\sigma^{\mathrm{train}}$ and tested on different $\sigma^{\mathrm{test}}$. Average PSNR calculated over BSD68 bsd. (c,d) Visual comparison of different networks tested on noise-level $\sigma^{\mathrm{test}}=50$. Details are better visible by zooming.
• Figure 4: (a) Performance of demosaicing CDLNets trained on $\sigma^{\mathrm{train}}=[01,20]$ and tested on different $\sigma^{\mathrm{test}}$. CDLNet and CDLNet-B are models with and without adaptive thresholds respectively. Average PSNR calculated over CBSD68 bsd. (b) Visual comparison of JDD CDLNet and JDD CDLNet-B tested on noise-level $\sigma^{\mathrm{test}}=30$. Details are better visible by zooming.
• Figure 5: Visual comparison of JDD CDLNet and state-of-the-art joint denoising and demosaicing methods on a crop of image 270 from the MIT moiré dataset Gharbi2016 with $\sigma=0$. Results from other methods are obtained from Liu_2020_CVPR.