Abstract
Recent years have witnessed the success of deep networks in compressed sensing (CS), which allows for a significant reduction in sampling cost. We propose PCNet, a practical and compact network for general image CS. PCNet features a novel collaborative sampling operator (COSO) consisting of deep conditional filtering and dual-branch fast sampling. This approach learns an implicit representation of a linear transformation matrix and performs adaptive local filtering. PCNet utilizes an enhanced proximal gradient descent algorithm-unrolled network for reconstruction, offering flexibility and interpretability. We also provide a deployment-oriented extraction scheme for single-pixel CS imaging systems. Experiments show superior reconstruction accuracy and generalization, especially for high-resolution images.
Introduction
Compressed Sensing (CS) enables the reconstruction of high-resolution images from a fraction of the standard measurements, offering significant savings in time, energy, and bandwidth. While deep learning has advanced CS reconstruction, the sampling process—capturing these measurements—remains a bottleneck. Existing deep CS methods often struggle with limited receptive fields (leading to blocking artifacts), a lack of interpretability in "black box" learned operators, and a disconnect from physical hardware like Digital Micro-mirror Devices (DMDs).
Methodology
PCNet addresses these challenges through a tripartite architecture comprising a Sampling Subnet, an Initialization Subnet, and a Recovery Subnet.
Collaborative Sampling Operator (COSO)
The core innovation is the COSO, which replaces traditional dense matrices with a two-step process:
- Deep Conditional Filtering: A lightweight 7-layer CNN adaptively smooths the input image based on the target sampling rate.
- Dual-branch Fast Sampling: The signal is split between a Discrete Cosine Transform (DCT) branch for low-frequency structures and a scrambled block-diagonal Gaussian matrix branch for global receptive fields at low computational cost.
Modernized Reconstruction and Deployment
The reconstruction backbone utilizes an unrolled Proximal Gradient Descent (PGD) algorithm, enhanced with Swin-Conv blocks and spatial downscaling. Crucially, PCNet includes a deployment-oriented extraction scheme. This algorithm allows the conversion of any linear sampling operator into an explicit matrix form, making it directly loadable onto hardware like DMDs.
Experimental Results
PCNet was evaluated on standard benchmarks (Set11, CBSD68) and high-resolution datasets (Test2K, Test4K).
- State-of-the-Art Performance: PCNet and its Transformer variant (PCT) outperform existing methods by 0.37–1.88 dB on standard benchmarks.
- High-Resolution Scalability: On 4K images, the performance gap widens to 3–6 dB over competitors, highlighting the importance of global sampling.
- Robustness: The model shows significant leads in quantized CS (1-bit or 8-bit) and improves self-supervised reconstruction by ~2 dB over traditional Gaussian sampling.
Conclusion
PCNet bridges the gap between theoretical deep CS and practical hardware deployment. By combining learned adaptive filtering with fixed mathematical transforms, it achieves superior performance and scalability. Future work will focus on physical optical bench implementations and extending the framework to specialized domains like medical MRI.
Summary updated