Reconstruction-free Cascaded Adaptive Compressive Sensing

Chenxi Qiu, Tao Yue, Xuemei HuCVPR 2024 · DOI

Abstract Adaptive Compressive Sensing (ACS) enhances signal acquisition by allocating sampling resources to information-dense regions. However, traditional multi-stage ACS methods suffer from a "reconstruction bottleneck," requiring slow intermediate image reconstructions to guide subsequent sampling. This paper introduces a reconstruction-free cascaded framework that utilizes a lightweight "ScoreNet" to allocate sampling rates directly from raw measurements. By bypassing intermediate reconstructions during inference, the proposed system achieves up to a 6x speedup over existing state-of-the-art methods while maintaining superior imaging fidelity.

Introduction

Compressive Sensing (CS) enables the capture of high-resolution signals at sub-Nyquist rates. Adaptive CS further optimizes this by non-uniformly distributing the sampling budget based on local scene complexity. Despite its potential, the computational overhead of intermediate reconstructions in multi-stage adaptive sampling has limited its practical application in real-time scenarios like medical imaging or high-speed photography. This work proposes an end-to-end pipeline that eliminates the need for reconstruction within the sampling loop, effectively removing the primary computational bottleneck.

Methodology

The proposed framework relies on several key innovations to enable efficient, reconstruction-free adaptive sampling:

  1. ScoreNet and Cascaded Framework: Instead of reconstructing pixels to determine complexity, a lightweight ScoreNet processes raw CS measurements to generate a complexity score for each image block. This score determines the additional sampling rate for the next stage.
  2. Differentiable Adaptive Sampling Module (ASM): To allow end-to-end training, the sampling allocation is modeled as a differentiable Linear Programming problem using a "perturbed optimizer" technique. This ensures that the sampling strategy is optimized directly to minimize final reconstruction error.
  3. MGSANet: A Multi-Grid Spatial-Attention Network is employed for the final reconstruction. Its parallel architecture is designed to maximize GPU utilization, offering high-speed inference compared to traditional sequential deep-unfolding networks.
  4. Out-of-loop Supervision: During training, intermediate reconstructions provide "fidelity supervision" to guide ScoreNet. Crucially, these steps are omitted during inference, maintaining the speed advantage.
The proposed cascaded adaptive CS framework (a) uses ScoreNet to generate sampling scores from measurements, while the Adaptive Sampling Module (c) handles differentiable rate allocation.

Experimental Results

The model was trained on BSDS500 and VOC2012 and evaluated on standard benchmarks including Set11 and Urban100.

  • Quantitative Performance: On the Set11 dataset at a 25% sampling rate, the proposed method achieved a PSNR of 36.62 dB, surpassing the previous state-of-the-art (AMS-Net) by 0.27 dB and non-adaptive methods by over 1.5 dB.
  • Inference Speed: The system demonstrated significant efficiency, with an inference time of 0.0307 seconds on Urban100—approximately 6x faster than the high-performance OCTUF+ model. This speedup is attributed to the reconstruction-free sampling loop and the parallelized MGSANet.
  • Visual Fidelity: Qualitative results show that the adaptive sampling effectively captures fine textures and high-frequency details in complex urban scenes that non-adaptive methods often blur or fail to resolve.
Quantitative comparison (PSNR/SSIM) and visual reconstruction results on Set11 and Urban100 datasets across various sampling rates.

Conclusion

By removing the reconstruction bottleneck, this work makes adaptive compressive sensing viable for time-critical applications. The use of raw measurements for sampling guidance, combined with a parallelized reconstruction network, sets a new benchmark for both speed and quality in the field. Future research will explore extending this framework to temporal and volumetric data, such as video and 3D imaging.

Summary updated

BibTeX
@inproceedings{qiu2024reconstruction,
  author = {Chenxi Qiu and Tao Yue and Xuemei Hu},
  title = {{Reconstruction-free Cascaded Adaptive Compressive Sensing}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2024},
  doi = {10.1109/cvpr52733.2024.00253},
}
Figure