Large-aperture optics play key roles in the laser beam transmission processes involved in high-power laser facilities. The microtargets (e.g., contaminants, dust, and particles) adhering to the optical surfaces would greatly affect the optical performance and, thus, need to be accurately detected for evaluating the cleanliness quality of optical components. However, due to the limit of camera resolution (the actual size of the target area on the optics surface represented by a single pixel ranges from 50 to 100 $\mu \text{m}$ ), it is of great challenge to accurately detect the sizes of tiny microtargets (roughly 20 $\mu \text{m}$ ). In this work, a novel subpixel size calibration method based on the regression model and the Mie scattering theory was proposed to precisely calculate the actual sizes of the tiny microtargets. The least-squares support vector machine (LSSVM) principle was applied to establish the area calibration model, and the random sampling consistency (RANSAC) algorithm was applied to optimize the selection of training samples and eliminate the outliers at the same time. The results showed that the relative diameter errors of about 90% of the detected microtargets were less than 30%, which is much better than that of the common pixel-level calibration method. The minimum detectable diameter of the microtargets with the proposed size calibration method can reach $15.8~\mu \text{m}$ , which is much smaller than the resolution ( $53.7~\mu \text{m}$ ) of commercial cameras. A similar high calibration accuracy can be achieved in different regions on the optical surfaces although the illumination conditions were different. The proposed subpixel size calibration method can be applied to detect the microtargets with dimensions as small as 20 $\mu \text{m}$ on the large-aperture reflector surfaces, which would greatly save the cost of detection equipment and improve the detection efficiency.