Enhancing stability in Zernike coefficient computation through deep learning for polygonal mirrors of electro-optical satellite payloads

In this study, we analyze the stability of Zernike coefficient computation using deep learning techniques and propose a new training method for deep learning model that can reliably output higher-order Zernike coefficients. Previous studies have shown that deep learning is a powerful tool for accurately deriving the Zernike coefficients of polygonal mirrors, but reliably extracting higher-order Zernike coefficients remains one of the challenges. To overcome these challenges, we present a new training method for the stability of deep learning model, enabling reliable high-order Zernike coefficient computation. The proposed deep learning model is designed based on the Network-in-Network concept, and a two-stage training process ensures that low-order and high-order Zernike coefficients are simultaneously reliably generated. Experimental results for performance evaluation show that the proposed deep learning model is effective in outputting stable and reliable higher-order Zernike coefficients, especially for polygonal mirrors.