Rapid focus position detection in satellite electro-optical payloads using fully convolutional networks

This study proposes a deep learning-based fast-focus detection network (FFDNet) to rapidly and accurately identify the optimal focus position throughout the entire assembly process of satellite electro-optical payloads. Built on a fully convolutional network architecture, FFDNet offers flexibility in predicting the focus positions, regardless of the input image size. It is characterized by rapid prediction speed, high accuracy, and robustness against external environmental noise. To address issues such as the excessive measurement time and sensitivity to environmental conditions found in existing algorithms, we trained FFDNet using knife-edge target images captured by a push-broom sensor. Experimental evaluations demonstrate that FFDNet outperforms existing algorithms in terms of both speed and accuracy and operates robustly under various temperature conditions and microvibration environments.