Large-size light guide plates (LGPs) with single edge lighting have the characteristics of uneven dot distribution, different defect sizes and shapes, complex background texture and so on. The traditional machine vision method of manually selecting features has insufficient generalization ability. In response, this study proposes a defect detection method based on improved YOLOv3 for large-size LGPs. Firstly, the improved multi-branch RFB module is introduced into the shallow feature layer of the network to increase the network receptive field, enrich the target semantic information and strengthen the ability of feature extraction. Secondly, the depth separable convolution is used to replace the standard convolution to reduce the size and calculation of the model. Furthermore, the K-means algorithm is improved to linearly scale the clustered anchor box so that it can be closer to the real box. Finally, a large number of experimental studies are carried out by using the defect pictures of large-size LGPs collected in a production site. The experimental results show that the average accuracy of the proposed detection algorithm is 98.92%. Compared with YOLOv3, this method has the average accuracy and F1 increased value by 8.55% and 10.76% respectively with a detection speed reaching 71.6 fps, which can meet the detection accuracy requirements of industrial production.