FSGAT-Net: Feature-space-based gated adaptive soft thresholding network for compressive sensing image reconstruction

期刊论文
发表日期：

2025年6月26日

查看全文

Shihao Li,
Qiang Guo,
Yong Wang,
Vladimir Tuz

摘要

In recent years, deep learning has made significant progress in the field of image compressed sensing (ICS), with many optimization-based network architectures proposed. These networks demonstrate excellent performance and interpretability by transforming traditional iterative reconstruction processes into deep unfolded networks (DUNs) and training them in an end-to-end manner. However, most unfolded networks still perform updates in the pixel space, without fully exploiting the feature information of the image, limiting the utilization of information flow. Moreover, existing DUNs based on $\ell _1$ $ℓ_{1}$ -regularization optimization typically use fixed thresholds for soft-thresholding operations, which fail to adapt to different image structures. To address these issues, we propose a Feature Space-based Gated Adaptive Soft Threshold Network (FSGAT-Net). This method enhances the feature representation in the reconstruction process through a feature space information enrichment module and introduces a gated adaptive soft-thresholding strategy. As a result, FSGAT-Net can adjust thresholds adaptively based on the image content, approximating the proximal mapping of $\ell _1$ $ℓ_{1}$ -regularization. Experimental results show that the proposed method performs excellently in both performance and robustness.