Research Area:  Machine Learning

Abstract:

In recent years, compressive sensing (CS) theory has garnered significant attention due to its advantages in high-resolution image processing. However, high-resolution images are susceptible to noise contamination prior to sampling, which leads to a noise folding phenomenon during CS denoising and reconstruction. This issue severely affects the visual quality of the reconstructed images. To address this issue, this paper proposes a Compressive Sensing Transformer Unfolding Network (CST-UNet) that leverages image degradation priors for high-quality image reconstruction. Specifically, we first design a Degradation Prior Gradient Descent (DPGD) module to learn noise degradation and guide adaptive gradient descent. Next, we develop a Dual-Path Transformer-CNN (DPTC) hybrid framework to capture both local and global contextual information, thereby mitigating block artifacts. Finally, we introduce inter-stage feature cross-attention (ISFCA) blocks to enhance information interaction between stages. Extensive experimental results demonstrate that the proposed CST-UNet achieves high visual quality for reconstructed images, even under conditions of noise pollution and low sampling rates.

Keywords:  

Author(s) Name:  Jie Zhang, Wenxiao Huang, Miaoxin Lu, Linwei Li, Yongpeng Shen, Yanfeng Wang & Jinsong Du

Publisher name:  Springer

DOI:  10.1007/s40747-025-01963-0

Volume Information:  Volume 11, article number 336, (2025)

Paper Link:   https://link.springer.com/article/10.1007/s40747-025-01963-0