During the process of apple blemish detection based on machine vision technology, due to the color similarity between stem/calyx and blemish, which greatly decreases the accuracy in apple detection, a method was proposed based on Decision Tree-Support vector machine (DT-SVM) to solve the challenge problem. Firstly, the single threshold method is used to remove the background. Then in the R channel, Connected Component Labeling method and Otsu method were employed to extract object regions (stem, calyx, blemish)，which were used to compute the color, texture and shape features. In the end, adopted the DT-SVM method to distinguish blemishes from the stem and calyx of apple images. By conducted on 600 apple images, the average accuracy of experiments was 97.7%. Compared to 1-V-1 SVM method and AdaBoost method, the DT-SVM method had a higher accuracy and less time-consuming, which could actually validate the effectiveness of the proposed method in recognizing the blemish of the apples.