Spatial compressive imaging deep learning framework using joint input of multi-frame measurements and degraded maps.

Area of Science:

• Computer Vision
• Image Reconstruction
• Machine Learning

Background:

• Traditional iterative methods for compressive imaging reconstruction are time-consuming.
• Deep learning approaches offer speed but lack flexibility and interpretability, especially at high compression ratios.

Purpose of the Study:

• To develop a novel deep learning network for efficient and flexible compressive imaging reconstruction.
• To address the limitations of existing deep learning models in handling high compression ratios and system variations.

Main Methods:

• Introduced Joinput-CiNet (joint input compressive imaging net), an end-to-end network.
• Integrated an imaging degradation model into the network input via a tailored encoding module.
• Trained and evaluated the network using diverse image datasets and infrared measurements.

Main Results:

• Joinput-CiNet demonstrated superior reconstruction performance compared to other networks.
• Achieved fast reconstruction speeds, particularly at low compression rates (e.g., 1:16, 1:64).
• The network effectively utilized prior knowledge of the imaging system for improved efficiency and performance.

Conclusions:

• The proposed Joinput-CiNet enhances training efficiency and reconstruction performance in compressive imaging.
• The joint input approach provides a flexible and interpretable deep learning solution for various compression ratios.
• This method offers a promising advancement for real-time and high-performance imaging applications.