Improved Faster R-CNN Based Surface Defect Detection Algorithm for Plates

Table 4

Effect of the number of deformable convolution layers on model performance.

Ablation experiments	Different strategies	Accuracy (%)	Recall rate (%)	Training time (h)	Average detection (s)
Number of deformable convolutional layers	3, 4, 5	92.16	85.38	22	0.61
	4, 5	96.64	89.23	18	0.52
	5 (√)	98.43	92.86	16	0.40