Bad pixel recognition and dark current modelling for CMOS with machine learning algorithm

Abstract

Complementary Metal Oxide Semiconductor (CMOS) detectors have attracted more and more interests as appropriate imaging instruments, because their performance is close to that of CCD with low power consumption and low noise level. Yangwang-1 is a commercial satellite in China that has two telescopes on board. An optical telescope and a ultra-violet telescope are installed in the satellite to observe near earth objects and bright stars in optical and ultra-violet band for space mining. Yangwang-1 uses CMOS as its camera for optical and UV-band. However, CMOS detectors have different response for different pixels and there is no independent cooling system for the CMOS in the Yangwang-1, which would make dark current control and modelling hard. In this paper, we propose an unsupervised dark current modelling and bad pixel recognition method for CMOS detectors in Yangwang-1 UV camera. Our method obtains several dark current calibration frames on the ground when the CMOS is in different temperatures. Then vectors of dark current in each pixel will be clustered with Gaussian Mixture Model to identify bad pixels and pixels that have the same trends in dark current and temperature relations. We would fit different dark-current and temperature relations for pixels that belong to the same cluster as templates for dark current estimation. Then we could quickly estimate dark current for real observation data with pixels that have no sources nearby. With estimated dark current and bad pixel mask, we find that our source detection pipeline could achieve higher accuracy for UV band observation images.