MMC-CS: Multi-branch multi-stage contrastive learning for self-supervised compressed sensing.

Area of Science:

• Computer Vision
• Deep Learning
• Signal Processing

Background:

• Deep neural networks (DNNs) show promise in image compressed sensing (ICS).
• Current DNN methods struggle with acquiring ground-truth data and underutilize measurement information.
• Optimization-inspired networks integrate optimization theory into DNNs for ICS.

Purpose of the Study:

• To propose a novel self-supervised deep learning framework for solving the inverse problem in ICS.
• To address the challenges of limited ground-truth data and underutilized measurements in DNN-based ICS.
• To develop an effective method for image reconstruction in the absence of labeled measurements.

Main Methods:

• A self-supervised deep learning framework leveraging measurement values via a multi-branch, multi-stage progressive cross-contrast structure.
• Design of a Multi-branch Multi-stage Cross-contrast CS (MMC-CS) end-to-end DNN, unfolding the Proximal Gradient Descent (PGD) algorithm.
• Integration of multi-scale co-optimization (image paths and convolutional feature paths) and wavelet convolution (WTConv) for enhanced reconstruction.

Main Results:

• The proposed method effectively learns image priors without ground-truth data.
• Achieved an average Peak Signal-to-Noise Ratio (PSNR) improvement of 0.3-1.6 dB over existing self-supervised approaches.
• Demonstrated strong potential to compete with state-of-the-art supervised methods in image reconstruction.

Conclusions:

• The novel self-supervised framework addresses key limitations in DNN-based ICS.
• The MMC-CS network offers superior image reconstruction performance compared to current self-supervised methods.
• The approach shows promise for practical applications requiring efficient and accurate image recovery from undersampled measurements.