Content-Aware Scalable Deep Compressed Sensing

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Image compressed sensing (CS) aims to reconstruct high-quality images from undersampled measurements.
• Existing CS methods often struggle with adaptive sampling rate allocation and fine-grained scalability.
• Efficient and robust reconstruction networks are needed to address these challenges.

Purpose of the Study:

• To develop a novel content-aware scalable network (CASNet) for efficient image compressed sensing (CS).
• To achieve adaptive sampling rate allocation, fine-granular scalability, and high-quality image reconstruction.
• To improve training convergence and network robustness for CS reconstruction.

Main Methods:

• A data-driven saliency detector and saliency-based block ratio aggregation (BRA) for sampling rate allocation.
• A unified learnable generating matrix for producing sampling matrices of any CS ratio.
• An optimization-inspired recovery subnet guided by saliency information and a multi-block training scheme.
• SVD-based initialization and random transformation enhancement (RTE) for accelerated training and robustness.

Main Results:

• CASNet demonstrates superior performance compared to existing CS networks.
• The proposed components and strategies show effective collaboration and mutual support.
• The network achieves adaptive sampling, scalable reconstruction, and high-quality results.

Conclusions:

• CASNet offers an efficient and effective solution for image compressed sensing.
• The content-aware and scalable approach significantly improves reconstruction quality.
• The developed strategies enhance training efficiency and network robustness.