In response to the low efficiency of artificial defect detection of the ceramic tile surface, this paper proposed a deep-learning YOLOv5 algorithm to detect the defects of the ceramic tile surface at the production line. For a start, figure cutting and segmentation, as well as data enhancement processing were performed against the data set. Then, data in the dataset was labeled through labelimg. In the end, the dataset was sent to the optimized YOLOv5 network model for iterative training, with the optimal weight used in the test. After comparison in the experiment, the detection accuracy of the YOLOv5 model is higher than that of the Faster RCNN, SSD and YOLOv4 model, with average accuracy of over 96% and an average detection time of 14 ms, So the method is advanced and practical in ceramic tile defect detection. The defects of ceramic tiles can be detected during the production process with this method.