10.2%。 A grading method of mangosteen at size and color was proposed based on computer vision. Taking the blue rollers as background, the mangosteen images were pre-segmented by double thresholds in RGB color space. Through morphological operation, contour trace and region fill, the whole mangosteen target was obtained. Lastly, the peduncle, pedicel and pericarp were identified by 2G-R-B and G factors. According to the centroid of peduncle and pedicel and the centroid of pericarp, the fruit posture was evaluated, and the diameter was extracted as the size grading criteria. Meanwhile, the difference of saturation and hue of pericarp area in HIS color space was the color grading criteria. A grading experiment was carried out for 200 mangosteens. The results indicated that the accuracy of diameter measurement is ±1.8mm, and the maximal scale of neighbor grade mixed by color is 10.2%.