Ground-Truth Free Meta-Learning for Deep Compressive Sampling

Xinran Qin, Yuhui Quan, Tongyao Pang, Hui Ji

Introduction

Compressive Sampling (CS) is a critical technique in fields like medical imaging (MRI, CT) and remote sensing, where it enables the reconstruction of high-resolution images from a limited set of measurements. However, CS is inherently an "under-determined" problem, as many potential images can correspond to the same sparse data.

Existing deep learning solutions generally fall into two categories: supervised learning, which requires expensive and often unavailable ground-truth (GT) images, and internal/self-supervised learning (like Deep Image Prior), which is computationally prohibitive because it requires training a new model for every single test image. This paper introduces MetaCS, a framework that bridges this gap by learning from external measurement datasets without ground-truth images, while maintaining high speed and accuracy.

The MetaCS Framework

MetaCS achieves its performance through four integrated technical pillars designed to optimize both training stability and adaptation speed:

  1. iSURE (Improved SURE) Loss: Traditional Stein’s Unbiased Risk Estimator (SURE) allows error estimation without ground truth but is computationally expensive due to complex second-order derivatives. MetaCS utilizes iSURE, which employs a noise-injection trick to simplify the mathematics, leading to faster and more stable backpropagation.
  2. Meta-Learning (MAML-based): The framework uses a modified Model-Agnostic Meta-Learning scheme to find a set of initial weights that are optimized for rapid adaptation. This allows the model to capture general image patterns from a broad collection of measurements.
  3. Null-Space Consistency (NC) Loss: To prevent the model from "hallucinating" details in the missing data regions (the null space), the NC loss ensures that test-time predictions remain consistent with the meta-learned prior.
  4. Bias-Tuning: For lightning-fast adaptation at test time, MetaCS only updates the network's bias parameters while keeping convolutional weights frozen. This targeted refinement acts similarly to thresholding in classical signal processing, allowing for significant quality improvements in seconds.

Experimental Results

The researchers evaluated MetaCS against medical (MRI150, ADNI) and natural image (Set11) benchmarks.

Visual Quality and Accuracy

MetaCS demonstrated visual reconstruction quality that is competitive with supervised models, despite never seeing a ground-truth image during training. It consistently outperformed other ground-truth-free methods like REI and ASGLD in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

Visual comparison of MetaCS against supervised and other unsupervised methods on MRI and natural image datasets.

The Impact of Meta-Learning

A key finding was the efficiency of the meta-learning stage. By starting from a meta-learned initialization, the model achieves higher PSNR values much faster than models trained from scratch or without the meta-learning prior. This advantage is consistent across different sampling rates and datasets.

PSNR vs. Iteration number showing the performance boost provided by the meta-learning stage across ADNI and Set11 datasets.

Speed and Robustness

MetaCS offers a massive speedup over traditional internal learning methods. For example, while some self-supervised methods take nearly five hours to process the Set11 dataset, MetaCS completes the task in less than one minute. Furthermore, the model proved robust to "unseen" measurement masks, handling variations in sampling rates better than traditional external learning approaches.

Conclusion

MetaCS represents a significant step forward in ground-truth-free image reconstruction. By combining the efficiency of meta-learning with the mathematical rigor of iSURE and bias-tuning, it provides a practical solution for high-stakes imaging domains where ground-truth data is scarce. Future work will focus on extending this framework to other ill-posed inverse problems, such as deblurring and super-resolution, and handling unknown noise distributions.

Summary updated

Abstract

Compressive sampling (CS) is an efficient technique for imaging. This paper proposes a ground-truth (GT) free meta-learning method for CS, which leverages both ex-ternal and internal deep learning for unsupervised high-quality image reconstruction. The proposed method first trains a deep neural network (NN) via external meta-learning using only CS measurements, and then efficiently adapts the trained model to a test sample for exploiting sample-specific internal characteristic for performance gain. The meta-learning and model adaptation are built on an improved Stein's unbiased risk estimator (iSURE) that provides efficient computation and effective guidance for accurate prediction in the range space of the adjoint of the measurement matrix. To improve the learning and adaption on the null space of the measurement matrix, a modi-fied model-agnostic meta-learning scheme and a null-space consistency loss are proposed. In addition, a bias tuning scheme for unrolling NNs is introduced for further acceler-ation of model adaption. Experimental results have demonstrated that the proposed GT-free method performs well and can even compete with supervised methods.
BibTeX
@inproceedings{qin2023ground,
  author = {Xinran Qin and Yuhui Quan and Tongyao Pang and Hui Ji},
  title = {{Ground-Truth Free Meta-Learning for Deep Compressive Sampling}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023},
  doi = {10.1109/cvpr52729.2023.00959},
}
Figure