Abstract: The intrinsic distortion of the fisheye lens in an all-sky camera can significantly alter the mapping between pixel coordinates and the true azimuth and zenith angles, making high-quality training star points essential for geometric correction prior to physical analysis. Because lens calibration models are sensitive to their initial parameters, traditional calibration methods based on manual annotation or rule-based matching often struggle to achieve both high efficiency and high accuracy. In this study, we propose an automated calibration method for all-sky cameras based on constellation recognition. The method first uses deep learning to automatically construct high-quality training star points to initialize the lens calibration model, and then iteratively improves the calibration accuracy by progressively expanding the training set. Experiments on observational data from an all-sky airglow imager at the Xinglong station of the Meridian Project show that the constellation recognition model achieves an mAP_50-95  of 0.985, and the resulting geometric calibration yields azimuth and zenith-angle correction errors within 2' , significantly outperforming a conventional distance-matching approach (within 45' ). This study provides an efficient and high-precision solution for automated calibration of all-sky cameras.