Hybrid model for compressive image recovery: Integrating ResNet-based denoiser into GAMP

• Published in: Signal processing 2020-08, Vol.173, p.107583, Article 107583
• Main Authors: Chen, Qun, Zhang, Haochuan
• Format: Article
• Language: English
• Subjects: GAMP; LD-GAMP; ResNet
• ISSN: 0165-1684; 1872-7557
• DOI: 10.1016/j.sigpro.2020.107583

•LD-GAMP blends a data-driven model, ResNet, into a hand-designed algorithm, GAMP.•An interpretable deep network is obtained by unfolding the iterative algorithm.•The network is insensitive to changes in measurement matrices.•The network has a small model size and is easier to train. Neural networks have witnessed great success in the recovery of compressive signals; however, these networks are largely unprincipled black boxes that are difficult to train, restricted to specific measurement matrices, and/or not applicable to non-linear processing, like quantization. To circumvent all these obstacles, the paper proposes a new network, LD-GAMP, that integrates a data-driven model, ResNet, into a hand-designed algorithm, GAMP, following an unrolling methodology that unfolds the iterative sparse-signal-recovery algorithm to form an interpretable deep network. Comparing to other state-of-the-art methods, the proposed algorithm is not only more capable but also less complex. On one hand, it is capable of recovering the quantized images with a high performance that exceeds both LD-AMP and GAMP by a notable gain. On the other, it has a model size that is only one half of the best competing model which resembles the LD-AMP. For this reason, it is more stable and even easier to train. Besides, the new method is insensitive to the change of measurement matrices during the training and the testing phases, owing to the generality of the GAMP algorithm. Also, it runs much faster than the competing non-learning based method as it avoids the time-consuming optimization process.