CSformer: Bridging Convolution and Transformer for Compressive Sensing.

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) excel at local feature extraction but have limitations in capturing global context.
• Transformer networks effectively model global dependencies using self-attention mechanisms.
• Integrating CNNs and transformers for compressed sensing (CS) remains an open research challenge.

Purpose of the Study:

• To propose CSformer, a novel hybrid framework for end-to-end compressive image sensing.
• To leverage the complementary strengths of CNNs and transformers for improved image reconstruction.
• To enhance the representation capacity by effectively combining local and global features.

Main Methods:

• Developed an adaptive sampling module with a learned sampling matrix for block-by-block image measurement.
• Designed a reconstruction stage featuring concurrent CNN and transformer stems for feature extraction and aggregation.
• Incorporated an initialization stem for efficient, learnable initial reconstruction.
• Utilized a progressive strategy and window-based transformer blocks to optimize computational complexity.

Main Results:

• CSformer demonstrates superior performance in compressive image sensing compared to state-of-the-art methods.
• The hybrid architecture effectively captures both fine-grained local and long-range global features.
• Experimental results validate the effectiveness of the proposed transformer-based approach on diverse datasets.

Conclusions:

• The proposed CSformer framework offers a powerful and effective solution for compressive sensing.
• Hybrid CNN-transformer architectures hold significant promise for advancing image reconstruction techniques.
• CSformer provides a new benchmark for performance in compressed sensing applications.