In the process of growth, harvest and transportation of carrots, it is inevitable that carrots appear some external defects. The elimination of defect carrots is an important link before carrot marketing. However, carrots mainly rely on manual grading nowadays, which has the inherent disadvantages of unstable grading standards, high labor consumption and high cost. In order to detect defective carrots quickly, accurately, and non-destructively, machine vision technology was introduced into carrot grading process to improve the classification accuracy and efficiency. Carrot external defects included green shoulder, bending, broken, furcation, and cracking. Different detection algorithms were proposed for different defects, since the different defects had different characteristics. The detection of green shoulder was realized by color difference between normal area and green shoulder area. In the HSV color space of carrot image, the threshold values of H, S, and V in region of green shoulder were determined by statistical method. Moreover, the recognition of bending, broken, and furcation were based on the shape difference between normal and defect carrots. The algorithm of convex hull, Hu moment invariants, and Harris corner detection methods were used to identify bending, broken, and furcation respectively. Furthermore, the detection of cracking was recognized by the difference texture of carrot. Sobel and Canny edge detection algorithm combined with morphologic operator to extract cracking region of carrot. The experimental results showed that the recognition accuracy of green shoulder, bending, broken, furcation, and cracking were 100%, 91.14%, 90.57%, 94.57%, and 95.45% respectively, and the overall recognition rate was 94.91%. The proposed defect recognition algorithm of carrot can provide algorithm reference for subsequent defect carrot online detection.