Abstract: Semi-mechanical method is often used for tangerine classification, which has low efficiency and uneven quality. This paper proposed a key technology for tangerine classification based on machine vision. The three channels of the original color image are converted into the gray image with only one channel by gray-scale, and the target area of the orange image is segmented by binarization, and the target area of the orange image is framed by the minimum external rectangle to realize the orange region recognition. Finally, the orange size is graded by calculating the smallest external rectangle. The experimental results show that the average accuracy of orange size classification based on machine vision is 96.7%, which can effectively meet the requirements of orange size classification.