The imaging characteristics of an image with light spots and the grey-level distribution model of an ideal light spot were analyzed. A fast and accurate algorithm to detect simultaneously multiple light spot centers in a complex imaging environment was proposed for the image with multi-scale light spots. As the centers of light spots would not be changed after blurring the spot image with Gaussian kernels, the image with massive multi-scale light spots was blurred firstly with multilevel Gaussian kernels to fast establish a Gaussian scale-space of the spot image. Then an efficient non-maximum suppression algorithm was applied to find local extremums in multiple scales and to determine the pixel level coordinates of the light spot centers in the scale-space preliminarily. Finally, combined with the neighboring pixels of these pixel level coordinates, sub-pixel accurate locations of the spot centers were obtained by local surface fitting. The validity of proposed algorithm was verified by simulation and experiments. The results for an image of 640 pixel ×480 pixel show that the processing time is 50 ms, average detection time for per thousand light spots is only 23 ms and the detection accuracy is 89% in many complex situations. Moreover, the algorithm is sensitive to low-light spots and can process the images with different scale spots in low contrast scenes, usually offering a low error rate and miss rate. Due to the high detection speed and good stabilitiy, the proposed algorithm performs well in real vision measurement systems.