Fig. 9 Comparison chart between model Loss curve and model mAP value
⑶ P-R curve
The pink, purple, green and gold curves in the figure represent the P-R curves of pit, wear, snowflake and crack respectively. The P-R curves of pit completely cover the P-R curves of crack, which indicates that the improved RetinaNet model has better prediction performance for pit than that of crack. Wear’s P-R curve intersects with Snowflake’s. Predictive performance is compared by equilibrium points. The P-R curve of snowflake has break-event Point (BEP) A, and the P-R curve of wear has BEP B. Precision at points A and B is equal to recall. Point B is larger than point A, indicating that the prediction performance of improved RetinaNet model for wear is better than that of snowflake. The order of defect prediction performance is pit, wear, snowflake and crack.