At present, the defects on the surface of the light guide plate are mainly detected by human eye, only a few manufacturers use the traditional image processing methods. In the imaging of high-resolution industrial cameras, the defects of the light guide plate are still extremely small, the characteristics of different defects are different, and the light guide points of the entire light guide plate are densely distributed and uneven,which leads that the traditional image processing detection methods require experienced visual experts to carry out a large number of feature extraction algorithm programming work and expensive code maintenance cost, low accuracy and poor stability. Therefore, a defect detection method based on deep learning semantic segmentation is proposed. This method can learn and extract the characteristics of the light guide plate defects by training the neural network to avoid the complicated feature extraction algorithm programming. First, the collected light guide plate defects are marked for making a sample set. Secondly, the pre-trained Pyramid Scene Parsing Network (PSPNet) is used to retrain the labeled samples using transfer learning. Further, the trained model is used to achieve detection of defects of the light guide plate. Since the separate deep learning semantic segmentation defect detection method usually cannot meet the industrial practical application requirements, it is necessary to combine the simple machine vision method to make a second judgment and screening of all suspected defect regions detected by the deep learning semantic segmentation method. The experimental results show that the detection rate of the three defects of bright spots, dark spots, and scratches is as high as 96%, which can basically meet the industrial testing requirements.