Coded aperture compressive temporal imaging via unsupervised lightweight local-global networks with geometric...

Area of Science:

• Optics and Photonics
• Computer Vision
• Signal Processing

Background:

• Coded aperture compressive temporal imaging (CACTI) uses compressive sensing (CS) to capture high-dimensional (HD) signals in a single snapshot.
• Deep learning excels at signal reconstruction but often requires large models and extensive training datasets, limiting practical applications.

Purpose of the Study:

• To develop a lightweight deep learning network for reconstructing HD signals from noisy, compressed measurements.
• To design an unsupervised reconstruction method that overcomes the limitations of traditional supervised deep learning approaches in optical imaging.

Main Methods:

• Proposed a lightweight convolutional neural network architecture designed to extract and fuse local and global features from compressed measurements.
• Developed unsupervised loss functions based on signal geometric properties to enhance network generalization for real optical systems.
• Implemented a novel block structure for feature extraction and fusion within a multi-layered network.

Main Results:

• The proposed lightweight network significantly reduces model size compared to existing methods.
• Achieved high performance in reconstructing dynamic scenes from compressed measurements.
• The unsupervised reconstruction network demonstrated performance comparable to its supervised counterpart.

Conclusions:

• The developed lightweight, unsupervised network offers an efficient and effective solution for HD signal reconstruction in CACTI systems.
• This approach broadens the applicability of deep learning in real-world optical imaging by reducing data and computational requirements.
• The unsupervised method provides a viable alternative to supervised learning, achieving strong generalization and reconstruction quality.