AMP-Net: Denoising based Deep Unfolding for Compressive Image Sensing (2004.10078v2)

Abstract: Most compressive sensing (CS) reconstruction methods can be divided into two categories, i.e. model-based methods and classical deep network methods. By unfolding the iterative optimization algorithm for model-based methods onto networks, deep unfolding methods have the good interpretation of model-based methods and the high speed of classical deep network methods. In this paper, to solve the visual image CS problem, we propose a deep unfolding model dubbed AMP-Net. Rather than learning regularization terms, it is established by unfolding the iterative denoising process of the well-known approximate message passing algorithm. Furthermore, AMP-Net integrates deblocking modules in order to eliminate the blocking artifacts that usually appear in CS of visual images. In addition, the sampling matrix is jointly trained with other network parameters to enhance the reconstruction performance. Experimental results show that the proposed AMP-Net has better reconstruction accuracy than other state-of-the-art methods with high reconstruction speed and a small number of network parameters.

• The paper presents AMP-Net, a deep unfolding model that enhances compressive image sensing by unfolding the AMP algorithm with integrated denoising and deblocking processes.
• The methodology leverages a jointly trained sampling matrix and modular reconstruction blocks to improve reconstruction accuracy, as evidenced by higher PSNR and SSIM metrics at low sampling rates.
• Experimental results on Set11 and BSDS500 demonstrate AMP-Net’s computational efficiency and ability to reduce blocking artifacts compared to traditional CS methods.

A Comprehensive Review of "AMP-Net: Denoising based Deep Unfolding for Compressive Image Sensing"

The paper "AMP-Net: Denoising based Deep Unfolding for Compressive Image Sensing" by Zhonghao Zhang et al. presents an intricate exploration into the development of a novel deep unfolding model termed AMP-Net, tailored explicitly for improving the performance of compressive sensing (CS) in visual images. This research capitalizes on the advantages of both model-based and deep network methods, synergizing them into a cohesive framework.

Approach and Methodology

AMP-Net is built upon the tenets of deep unfolding, where the authors introduce a novel approach by unfolding the approximate message passing (AMP) algorithm. Unlike traditional methods that focus on learning explicit regularizers, AMP-Net focuses on unfolding the denoising process of AMP, catering to both reconstruction accuracy and speed. This method distinguishes itself by incorporating a deblocking module, addressing typical blocking artifacts in CS of visual images. Additionally, it leverages a jointly trained sampling matrix with network parameters, which significantly enhances reconstruction accuracy and performance.

The model operates in two distinct segments: a sampling model and a reconstruction model. The latter is explicitly designed with an initialization module and multiple reconstruction modules, the latter of which integrates both denoising and deblocking functions. This modular design allows AMP-Net to mimic the iterative refinement process endemic to the AMP algorithm, while introducing a learnable structure that adapts to the characteristics of the image signal in the learning process.

Experimental Results

Significant empirical assessments underscore the efficacy of AMP-Net, demonstrating its superior performance against prevalent CS methods. Tested on the Set11 dataset and BSDS500, AMP-Net consistently delivered robust performance across various compression ratios. The results are quantified using PSNR and SSIM metrics, where AMP-Net exhibited noteworthy gains especially at lower sampling rates, achieving improved reconstruction accuracy and reduced computational complexity compared to state-of-the-art alternatives.

A key highlight of AMP-Net is its computational efficiency, attributed to its small number of network parameters, which facilitates rapid image reconstruction—a crucial advantage in real-world applications requiring swift processing. Additionally, the trainable sampling matrix introduced in AMP-Net not only optimizes the data sampling process but also integrates into existing generative networks, broadening its utility.

Implications and Future Directions

The implications of AMP-Net extend beyond its superior performance metrics. By fostering a reconciliation between model-based interpretative robustness and the flexibility of deep networks, AMP-Net sets a precedence for future architectures in the domain of image reconstruction. The joint training of sampling matrices and network parameters introduces a paradigm where the data acquisition process is naturally coupled with reconstruction, potentially catalyzing advancements in adaptive imaging systems and edge computing.

The exploration of AMP-Net propels the discourse on deep learning unfolding methodologies, suggesting avenues for integrating complex signal structures and multifaceted priors in image signal processing. Future research could explore the scalability of AMP-Net in multi-dimensional signals and its adaptability in hybrid models that incorporate variational inference frameworks or generative models, which could potentially unlock new capabilities in dynamic and real-time imaging applications.

In conclusion, "AMP-Net: Denoising based Deep Unfolding for Compressive Image Sensing" provides a comprehensive model that amalgamates interpretative clarity with computational efficiency in compressive sensing, proffering a significant leap forward in visual image processing aided by deep unfolding techniques.