Edge detection for optical synthetic apertures based on conditional generative adversarial networks.

Area of Science:

• Optical Engineering
• Image Processing
• Machine Learning

Background:

• Detecting interference fringes in optical synthetic apertures is crucial for sub-mirror alignment (piston, tip, tilt).
• Existing edge detection methods struggle with complex fringe shapes, gaps, and varying optimal thresholds, often losing boundary information.
• Traditional algorithms fail to restore lost boundary data, impacting alignment accuracy.

Purpose of the Study:

• To propose a novel, robust method for detecting the edge of synthetic aperture fringe images.
• To overcome limitations of traditional edge detection algorithms in optical synthetic aperture systems.
• To improve the accuracy and reliability of sub-mirror alignment by enhancing fringe detection.

Main Methods:

• Utilized conditional generative adversarial networks (cGANs) to address challenges in designing task-specific loss functions.
• Trained the cGAN model on a dataset of over 8000 real and simulated fringe images.
• Developed a global decision-making process to detect and fill missing boundary information.

Main Results:

• The proposed cGAN-based method achieved a false detection rate of 0.2, significantly outperforming the Canny algorithm's rate of 0.56.
• Successfully detected fringe edges in systems with varied sub-mirror shapes and increasing numbers.
• Demonstrated the ability to globally infer and fill lost boundary information with a maximum error of two pixels.

Conclusions:

• Conditional generative adversarial networks offer a superior solution for edge detection in complex optical synthetic aperture fringe images.
• The novel method enhances sub-mirror alignment by providing more accurate and complete fringe boundary information.
• This approach represents a significant advancement for optical synthetic aperture systems, particularly those with dynamic and complex configurations.