AutoBCS: Block-Based Image Compressive Sensing With Data-Driven Acquisition and Noniterative Reconstruction

Yang Gao, Hongping Gan, Yang Gao 0030, Chunyi Liu, Haiwei Chen, Tao Zhang, Feng LiuIEEE Trans. Cybernetics 2023 · DOI

TL;DR

This paper introduces AutoBCS, a novel deep learning framework for compressive sensing (CS) that unifies data-driven image acquisition and non-iterative reconstruction. By learning a specialized sensing matrix and employing an efficient reconstruction network, AutoBCS significantly improves image quality and reconstruction speed compared to traditional and existing deep learning CS methods.

What Problem Does This Paper Address?

Compressive Sensing (CS) is a powerful technique for acquiring and reconstructing signals at sub-Nyquist rates, offering energy efficiency and reduced data storage. However, traditional CS methods, particularly Block Compressive Sensing (BCS) for images, face several challenges:

  1. Non-Data-Driven Acquisition: The sparsifying domains and sensing matrices are often handcrafted and not tailored to the specific characteristics of the images being captured. This leads to a loss of valuable prior image information.
  2. High Computational Complexity: Reconstructing images from compressed measurements often involves solving complex optimization problems, which can be computationally expensive and time-consuming, especially for high-resolution images.
  3. Blocking Artifacts: Block-based approaches, while simplifying reconstruction, can introduce noticeable blocking artifacts at the boundaries of image blocks.

These limitations hinder the practical application of CS in resource-constrained environments like wireless sensor networks.

Methods

AutoBCS tackles these issues by proposing a unified, end-to-end deep learning framework with two main components:

  1. Data-Driven Image Acquisition Module:
    • Learning-Based Sensing Matrix (LSM): Instead of using predefined sensing matrices, AutoBCS learns a specialized sensing matrix (LSM) from training data. This LSM is designed to capture image-specific features and relationships among subblocks, leading to more efficient data acquisition.
  2. Convolutional Sampling: The acquisition process is implemented using a convolutional layer, where the learned weights of the filters form the LSM. This approach is proven to satisfy theoretical CS requirements like the Restricted Isometry Property (RIP).
  3. Noniterative Image Reconstruction Module:
    • Initial Reconstruction Subnetwork: This subnetwork takes the compressed measurements and produces an initial, albeit potentially artifact-ridden, reconstruction. It uses a convolutional layer with learned weights, replacing traditional linear MMSE estimation.
  4. Octave Reconstruction Subnetwork: This novel subnetwork, based on a modified U-Net architecture incorporating octave convolutions and octave-transposed convolutions, refines the initial reconstruction. Octave convolutions process features at different spatial resolutions, effectively reducing computational cost and improving the representation of multiscale information. This subnetwork is crucial for eliminating blocking artifacts and enhancing image quality.

The entire AutoBCS framework is trained end-to-end, jointly optimizing both the acquisition and reconstruction modules for maximum performance.

Datasets

  • Training: The model is trained on 400 images from the BSD500 database, with data augmentation techniques applied.
  • Testing: The performance is evaluated on several standard benchmark datasets: Set5, Set11, Set14, BSD68, and BSD100.

Key Insights and Ideas

  • Unifying Acquisition and Reconstruction: The most significant insight is the integration of a learned sensing matrix directly into the deep learning pipeline, bridging the gap between signal acquisition and reconstruction.
  • Data-Driven Sensing: Learning the sensing matrix from data, rather than relying on handcrafted matrices, allows the system to adapt to image characteristics, leading to improved sampling efficiency.
  • Octave Convolutions for Refinement: The use of octave convolutions in the reconstruction subnetwork is a key innovation. It allows for efficient processing of multi-resolution features, crucial for denoising and artifact removal without significantly increasing computational load.
  • Noniterative Reconstruction: The framework achieves high-quality reconstruction without iterative optimization, drastically reducing reconstruction time.

Key Results

  • Superior Image Quality: AutoBCS consistently outperforms traditional BCS methods (IRLS, D-SPL, MH-SPL, WaPT, GBSR) and other deep learning-based methods (SDA-net, ReconNet, DR2-net, ISTA-net, CS-net) in terms of SSIM and PSNR metrics across various datasets and sampling rates. For example, it achieves significant PSNR gains over traditional methods.
  • Drastic Speedup: AutoBCS offers a substantial reduction in reconstruction time, being orders of magnitude faster than traditional BCS algorithms and comparable to other DL-based methods, while achieving better accuracy.
  • Robustness to Noise: The proposed method demonstrates good performance even in noisy conditions, outperforming traditional methods with significantly smaller PSNR degradation.
  • Effective LSM: The learned LSM, when used with traditional BCS frameworks, also shows improved sampling efficiency compared to standard sensing matrices.

Limitations and Future Work

  • Fixed Block Size: A limitation noted is the fixed sampling block size, which might require zero-padding for images not perfectly divisible by this size. The authors suggest investigating variational sampling matrix sizes for different regions in future work.
  • Understanding LSM Superiority: Further investigation is planned to understand why the learned LSM outperforms random matrices, even though both satisfy theoretical requirements.
  • Noisy Reconstruction: While robust to noise, the authors plan to explore stronger, more robust modified DL strategies for noisy image reconstruction scenarios.
  • Other Inverse Problems: The authors suggest that the framework could be extended to other image inverse problems like inpainting and deconvolution.

Summary updated

BibTeX
@article{gao2023autobcs,
  author = {Yang Gao and Hongping Gan and Yang Gao and Chunyi Liu and Haiwei Chen and Tao Zhang and Feng Liu},
  title = {{AutoBCS: Block-Based Image Compressive Sensing With Data-Driven Acquisition and Noniterative Reconstruction}},
  journal = {IEEE Trans. Cybernetics},
  year = {2023},
  doi = {10.1109/tcyb.2021.3127657},
}
Figure