TransCS: A Transformer-Based Hybrid Architecture for Image Compressed Sensing.

Area of Science:

• Signal Processing
• Computer Vision
• Machine Learning

Background:

• Compressed sensing (CS) is crucial for image acquisition and reconstruction.
• Reconstructing images accurately at low sampling rates remains a significant challenge.

Purpose of the Study:

• To propose a novel Transformer-based hybrid architecture (TransCS) for high-quality image compressed sensing.
• To improve image reconstruction accuracy, especially under low sampling conditions.

Main Methods:

• TransCS utilizes a trainable sensing matrix in the sampling module to learn image structural information.
• A customized iterative shrinkage-thresholding algorithm (ISTA)-based Transformer backbone models global dependencies among image subblocks.
• An auxiliary convolutional neural network (CNN) captures local image features.

Main Results:

• TransCS demonstrates superior reconstruction quality compared to state-of-the-art methods on benchmark datasets.
• The proposed method exhibits enhanced noise robustness.
• Experimental results validate the effectiveness of the hybrid Transformer-CNN architecture.

Conclusions:

• The TransCS architecture effectively addresses the challenge of image reconstruction in compressed sensing.
• The integration of Transformer and CNN components leads to high-performance image CS.
• TransCS offers a promising solution for accurate image reconstruction with reduced sampling rates.