Detection of Surface Defects in Lightweight Insulators Using Improved YOLOv5

Table 1. Software and hardware platform configuration parameters
View table
Table 1. Software and hardware platform configuration parameters
Environment Configure
GPU NVIDIA GeForce RTX3050
CPU Intel Core i5-11400H@2.7GHz
operating system Windows 10
framework PyTorch 1.9.0
CUDA cuda 11.1

Table 2. Comparison of different backbone networks
View table
Table 2. Comparison of different backbone networks
Model R_mAP@0.5 /% Model size /MB FPS
YOLOv5s 90.43 14.1 98
+Ghost 88.71 10.2 108
+C3TR 87.46 20.5 82
+ConvNeXt 93.27 13.6 105

Table 3. Comparison of different attention mechanisms
View table
Table 3. Comparison of different attention mechanisms
Method R_mAP@0.5 /% Model size /MB FPS
Baseline 90.43 14.1 98
+SE 90.31 15.3 98
+CBAM 92.26 13.8 103
+ECA 91.75 13.8 99
+CA 93.48 13.2 106

Table 4. Comparison of pruning rates
View table
Table 4. Comparison of pruning rates
Pruning rate /% P /% R /% R_mAP@0.5 /% Model size /MB
40 93.71 90.19 94.33 6.8
50 92.56 91.71 93.84 5.4
60 89.82 85.94 89.71 4.5
70 82.59 79.24 81.37 3.3

Table 5. Lightweight network comparison
View table
Table 5. Lightweight network comparison
Model R_mAP@0.5 /% Model size /MB FPS
YOLOv3-tiny 84.37 36.2 69
YOLOv4-tiny 86.93 23.7 75
YOLOX-nano 89.71 24.8 71
Improved YOLO v5 93.84 5.4 166

Table 6. Comparison of experimental results of different algorithms

View table

Table 6. Comparison of experimental results of different algorithms

Model	P /%	R /%	R_mAP@0.5 /%	R_mAP@0.5∶0.95 /%	Model size /MB	FPS
SSD	80.97	76.41	81.45	58.39	176.4	55
Faster R-CNN	88.12	84.69	88.06	67.59	230.8	18
YOLOv3	85.68	83.58	87.67	65.82	67.9	62
YOLOv5	89.53	89.18	90.43	90.43	14.1	98
Improved YOLOv5	92.56	91.71	93.84	73.61	5.4	166

Table 7. Ablation experiment
View table
Table 7. Ablation experiment
YOLOv5 ConvNeXt block CA Prune P /% R /% R_mAP@0.5 /% FPS
√ 89.53 89.18 90.43 98
√ √ 90.65 90.42 93.27 105
√ √ 89.79 90.23 93.48 101
√ √ √ 94.12 92.86 95.25 118
√ √ √ √ 92.56 91.71 93.84 166

Tools

Get Citation

Copy Citation Text

Yu Guo, Meiling Ma, Dalin Li. Detection of Surface Defects in Lightweight Insulators Using Improved YOLOv5[J]. Laser & Optoelectronics Progress, 2023, 60(24): 2412007

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites