Table 2.

Comparison of some classical defect detection models. mAP represents the average detection accuracy of the model for the PASCAL VOC 2007 dataset and the PASCAL VOC 2012 dataset.

Model	Improvement	Insufficient	mAP
R-CNN	mAP greatly improved; introduced RP+CNN.	The training steps are tedious, and the training speed is slow.	58.50% 62.00%
Fast R-CNN	mAP has been improved, and training time has been shortened.	Unable to meet real-time requirements, and unable to detect end-to-end.	70.00% 66.00%
Faster R-CNN	Both accuracy and speed are improved, thus enabling end-to-end detection.	Real-time detection cannot be achieved, and the calculation is large.	73.20% 70.40%
YOLOv4	Large residual blocks, SPP, and PANNnet are introduced, and Mish activation functions are used.	Small target detection is poor, and the recall rate is relatively low.	85.46% 86.68%
YOLOv5	Adaptive anchor calculation, focus structure, and CSP structure are introduced, and GIOU_Loss is adopted.	Small target recognition is unstable and requires massive training data.	89.10% 92.68%
SSD	Multiscale detection, default anchor, data enhancement.	Needs artificial setting box value, and small target recognition effect is poor.	76.90% 77.91%