An Image-Based Deep Learning Approach with Improved DETR for Power Line Insulator Defect Detection

Table 2

Performance of contrastive models on CPLID.

Model	Backbone				Speed (ms)
RetinaNet	ResNet101	0.992	0.953	0.762	80.6
RetinaNet	ResNet50	0.993	0.956	0.777	60.2
YOLOv5x	CSPDarknet53	0.995	0.960	0.794	26.6
YOLOv5l	CSPDarknet53	0.995	0.969	0.795	14.2
YOLOv5m	CSPDarknet53	0.995	0.960	0.795	8.2
TOOD	ResNet101	1.000	0.969	0.764	86.2
TOOD	ResNet50	1.000	0.964	0.787	65.7
Sparse R-CNN	ResNet101	1.000	0.981	0.758	86.9
Sparse R-CNN	ResNet50	1.000	0.992	0.785	67.1
DETR	ResNet101	1.000	0.965	0.752	63.3
DETR	ResNet50	1.000	0.985	0.753	41.3
DETR-Focal	ResNet101	1.000	0.974	0.769	63.3
DETR-Focal	ResNet50	1.000	1.000	0.797	41.3