To improve the precision, update rate and the anti-noise ability of star sensors, this study was focused on the extraction of star images. Firstly, the criterion to confirm whether a pixel attributes to the same star coverage with the local maximum pixel was established according to the gray distributing features of star images. Then, several kinds of methods for image extraction were introduced, such as the block division of imaging array, and the prediction of background noise. Finally, by regarding the peak value pixel as the origin of the circular region growing, the local region growing criterion was set up successfully based on the gray distributing features of star images. The simulation conducted in the case of no noises shows that all of the simulated star images in the reference star map can be extracted successfully, and the sub-pixel location accuracy is 0.028 2 pixel by using centroid method. Moreover, under upmost adverse condition with a Gaussian noise mean value of 20 and standard variance as high as 2.5, the success extraction rate can still reach 86.11% with a decreased centroiding accuracy of 0.219 6 pixel. The test about a faint star image in a star map with poor uniformity and a low SNR of 4.9 dB also proves the excellent detection ability of the proposed method, and it shows advantages of good real-time property and high correct extracting rate for star images.