A global centralized magnetic flux leakage small defect detection network

Abstract To solve the problem that magnetic-flux-leakage (MFL) small defects are difficult to accurately detect by machine learning methods, a global centralized magnetic flux leakage small defect detection network (RCFPNet) is proposed. RCFPNet consists of simulation data enhancement, improved feature extraction (backbone), an improved centralized feature pyramid (CFP) and a detection head network. The MFL defect data of various scales and shapes are simulated by ANSYS simulation software and superimposed with the actual detected MFL defects to expand the dataset. The Repvgg module is used to replace the 3*3 convolution of the backbone to improve the detection speed. An improved spatially explicit vision center scheme (EVC) and a global centralized regulation rule (GCR) for feature fusion networks are proposed for feature fusion networks. RCFPNet is based on an improvement of the YOLOv5 network. Experiments have proven that RCFPNet has improved detection speed and accuracy and has achieved good results in the detection of magnetic leakage small defects. Experiments show that when the IOU = 0.5, the accuracy rate of this algorithm is 96.1%, and the reasoning time is 8.9 ms.