An End-to-End Optical System Optimization Approach: Optimizing A Wide Field Telescope for Deep Learning Based Celestial Objects Detection Algorithms
Wennan Xiang, Peng Jia, Zheng Li, Jiameng Lv · 2024
Object detection stands as a crucial application for optical systems. However, the conventional design of optical systems for object detection has historically fixated on specific imaging quality evaluation criteria, a practice that may not seamlessly align with the results obtained by object detection algorithms. This mismatch is particularly conspicuous in the intricate relationship between the performance of recently developed deep learning based celestial object detection algorithms and the images captured by wide field telescopes. In response to this challenge, we propose a novel approach that integrates the optical system simulator and the deep learning based object detection algorithm within an optimizer framework. The deep learning based object detection algorithm is manually designed and trained using specific datasets. Simultaneously, the optical system simulator has the capability to generate a large amount of images based on the optical design, hardware specifics of the camera, and the observation mode. The results provided by the object detection algorithm, operating on simulated images, serve as the benchmark for evaluating the performance during the optimization of the optical system. To validate our approach, we optimize a primary focus telescope for a celestial object observation scenario, employing a Faster R-CNN based celestial object detection algorithm. The results show a noticeable enhancement in both detection efficiency and accuracy. Thus, our approach emerges as an innovative method for optimizing different optical imaging systems, in tandem with their respective deep learning based algorithms, tailored for various applications.