As the core raw material of the silk industry, the quality of cocoon directly determines the quality of the final silk products. The type of defects on the surface of cocoon is a key index to judge the quality of cocoon. In this paper, four types of cocoon defect data sets are constructed by adding corresponding labels to cocoon images through Labelimg. Then, the YOLOv5m model and YOLOv5s model are trained by using the constructed cocoon defect data set, and the training results are compared in detail. After comparative analysis, the results show that the YOLOv5 m model has good results in various evaluation models such as accuracy, mAP and F1 score. Finally, the trained two models are deployed on the experimental platform, and under the same experimental detection conditions, the two models are used to detect the same number of cocoons. The test results show that the detection accuracy of YOLOv5m model is 83.1 %, while that of YOLOv5s model is 70.9 %. The experiment also provides a practical basis for deep learning technology in cocoon defect detection.