Abstract The star sensor distortion calibration method mainly adopts the fitting method, which is limited by the number of calibration points and errors in the engineering application. The traditional least square fitting method or the toolbox fitting method will produce over-fitting phenomenon in the fitting process, resulting in a decrease in the accuracy of the star sensor's distortion calibration. This paper proposes a high-precision distortion calibration method based on the Dropout method. This method first networkizes the high-order surface distortion model of the star sensor, and then constructs the distortion model of the star sensor with part of the convolutional layer hidden. Recently, supervised learning is performed to complete the calibration of star sensor distortion model. The test results show that the use of the star sensor calibration method based on the Dropout method can effectively improve the training accuracy of the star sensor. Compared with the fitting results of the high-precision toolbox, the distortion calibration residual is increased from 2.02″ to 1.12″.
2021, Vol. 4 
