Abstract: A self-calibration method based on spatial high frequency energy was proposed in view of the geometric artifacts caused by misaligned geometry in cone-beam CT. An optimization model was constructed based on spatial high frequency energy of the reconstructed image. Part of the parameters were directly extracted from the projection image to diminish the search scope, and the optimal solution of geometry parameters was achieved by NM-simplex method that makes the spatial high frequency energy of CT image maximize. To improve the speed of the algorithm, GPU was used to accelerate the process of image reconstruction to reduce the reconstruction time. Without reprocessing the reconstructed image, the proposed method has less computing complexity compared with the existing ones. The experiment results show that the presented method has significant effect in geometric artifact calibration of the reconstructed images with high accuracy and the maximum relative error is less than 5%. It can notably reduce iterations without precision reduction, and algorithm execution efficiency has been raised 18.8%.