An Efficient and Accurate Asteroid Inversion Algorithm

This paper mainly describes the Adam method under the problem of retrieving asteroid rotation and 3D model from radar 2D images. We improve on the traditional method by introducing computational graph and back propagation algorithm in deep learning to improve the operation speed and calculation accuracy of the gradient, and adopt Adam optimizer to accelerate the convergence speed and avoid local minimum. We compare this method with the existing Gauss-Newton method, Levenberg-Marquardt method, and conclude that the Adam method is better than the Gauss-Newton method and Levenberg-Marquardt method in convergence speed, two-dimensional image fitting accuracy, three-dimensional model fitting accuracy.