The noises caused by star sensor electric circuit and the complex working environment faced by the star sensor when being applied in spacecraft both have serious impact on the quality of the star image, leading to the decrease of the star centroid extraction precision. Regarding the issues above, a star centroid extraction method, which has high precision and good anti-noise performance, is presented. Firstly, a multi-window sampling adaptive threshold segmentation method was used to reduce the noise of star image. Secondly, an improved gray cross projection method was proposed for star point coarse extraction. Finally, the local region growing method and the Gaussian surface fitting were used for star centroid fine extraction. The improved gray cross projection method can solve the problems caused by star point overlap so as to quickly and accurately find the rough area of star point and provide reliable “seed” points for the local region growing method. The local region growing method can fully reserve the edge of star point, which is beneficial to improve the star centroid extraction precision. Simulation results show that the star centroid extraction accuracy of the proposed method reached 0.075 5 pixel under the simulation condition with a noise standard deviation of 2. Compared with the traditional methods such as the scanning method, the star centroid extraction accuracy of the proposed method is greatly improved with good anti-interference performance, which has significant reference value for improving the accuracy of star sensor measurement under strong noise interference conditions.