Deep-Learned Regularization and Proximal Operator for Image Compressive Sensing.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Deep learning methods for image compressive sensing (CS) struggle with arbitrary sampling ratios and lack explicit learned regularization.
• Existing CS reconstruction approaches often lack flexibility and explicit deep-learned regularization terms.

Purpose of the Study:

• To address the limitations of current deep learning-based CS reconstruction methods.
• To develop a novel approach combining deep-learned regularization and proximal operators for improved CS image reconstruction.

Main Methods:

• Introduced a deep-learned regularization term using a residual-regressive network to measure image corruption and identify subspace.
• Developed a proximal operator with a dilated residual channel attention network to map distorted images to a clean image set.
• Employed an adaptive proximal selection strategy and a self-ensemble strategy for enhanced reconstruction performance.

Main Results:

• The proposed method achieves superior accurate reconstruction, with a PSNR gain over 1 dB compared to competing approaches.
• Demonstrated state-of-the-art performance in image CS reconstruction.
• State evolution analysis confirmed the effectiveness of the designed networks.

Conclusions:

• The novel deep learning framework effectively solves the CS reconstruction problem by integrating learned regularization and proximal operators.
• The method offers improved accuracy, flexibility, and state-of-the-art performance in image compressive sensing.