Abstract: In order to improve the carrot automatic grading system based on computer vision, which refers to the grading standard of purchases and sales of carrots, this article proposed a detection method of carrot with green-shoulder, fibrous roots and surface cracks which impact the carrot appearance grading greatly. Five hundred and twenty carrots were selected randomly as testing samples, and their photos were taken by camera for the next step of processing and research. To detect the fibrous roots, the skeleton was extracted from the binary image after necessary image preprocessing and binariazation. Based on the fact that normal carrots have 5 end points of skeleton, fibrous roots are detected by computing the number of the skeleton end points. The number of fibrous roots is computed by subtracting 5 from the number of skeleton end points and then divided by 2, which is taken as the measure of whether there are fibrous roots on the carrot image. Green-shoulders of carrots can be distinguished by the green color which is very obvious on the R component image, so it is detected by thresholding the R component image of the carrot and computing the area of the green-shoulder region. The ratio of the area of the green-shoulder region and whole carrot region is defined as green-shoulder ratio to measure whether there is green-shoulder on the carrot. As for surface cracks, surface cracks are more obvious on S component images than on the other components, so they are detected by region marking on the S component image and computing the area of surface cracks. The ratio of the area of cracks and the whole carrot is defined as the degree of surface cracking to measure whether there are cracks on the carrot surface. For each carrot image, the three quantitative criteria mentioned above are computed for statistics and analysis, and then the detection accuracy rates of the three criteria are tested. The result show that the detection accuracy rate for the green-shoulder, the fibrous roots and carrot surface cracking defects detection are 97.5%, 81.8%, and 92.3%, respectively. It is showed from the result that the detection algorithms of green-shoulder and fibrous roots have higher accuracy rates than that of surface cracks and have reached above 92%. However, the detection results of cracks are not ideal because the color of cracks area is similar to that of the normal surface and are not obvious on the two-dimensional image. Finally, the detection rates of the three criteria are calculated synthetically. Result shows that the overall accuracy rate can reach 91.3%, which can meet the need of defect detection. The method proposed in this paper has positive significance for the research of carrot appearance quality grading system and product line.