Lightweight Guidance Sampling and Deep Refinement Reconstruction Network for Adaptive Compressive Sensing
Zhaoxin Cai, Yunzhou Zhang, Haoyue Bai, Lu Wang, Tengda Zhang, Sizhan Wang, Shibo Zhang
AI summary
Problem
Existing multi-stage adaptive compressive sensing methods fail to capture global structural information for sampling guidance and suffer from tight coupling between sampling and reconstruction, causing biased allocation and suboptimal recovery.
Approach
The method decouples a lightweight guidance network that uses gradient-fused cross-attention to estimate sparsity and allocate sampling resources, from a deep refinement network that employs decoder dense feedback and multi-branch attention for high-fidelity reconstruction.
Key results
- Decouples sampling guidance and deep refinement for targeted optimization
- Introduces gradient-fused cross-attention to compensate for global content bias in undersampled regions
- Designs decoder dense feedback and multi-branch attention to bridge feature gaps and enhance detail restoration
- Achieves state-of-the-art PSNR and SSIM on BSD68, Set11, Urban100, and DIV2K datasets, particularly at 1% and 4% sampling rates
Why it matters
Enables more efficient and accurate image reconstruction in resource-constrained sensing applications like MRI and single-pixel imaging.
Abstract
Adaptive Compressive Sensing (ACS) has at- tracted increasing attention for its ability to progressively improve image reconstruction quality by dynamically adjust- ing sampling allocation. Multi-stage sampling is a promising strategy that leverages intermediate reconstructions to guide sampling without relying on image prior information. How- ever, existing multi-stage methods often struggle to capture global structural information, resulting in biased sampling and suboptimal performance. Furthermore, the strong dependency between intermediate reconstruction for sampling guidance and the final reconstruction can hinder targeted optimization. To address these issues, we propose LGDR-Net, a Lightweight Guidance Sampling and Deep Refinement Reconstruction Net- work. Specifically, the Gradient-Fused Cross-Attention (GFCA) module, embedded within a lightweight guidance network, leverages globally fused information to compensate for in- complete content during multi-stage sampling. Then, sampling resource allocation is driven by inter-stage reconstruction differ- ences, effectively exploiting image sparsity information. Finally, the Deep Refinement Network incorporates a Decoder Dense Feedback Mechanism (DDFM) to reduce cross-layer structural bias and a Multi-Branch Attention Fusion (MBAF) module for improved fine-texture representation. Extensive experiments demonstrate that our proposed LGDR-Net outperforms state- of-the-art methods, achieving an excellent trade-off between computational cost and reconstruction quality.