Defect Detection Algorithm of Infrared Insulator Image Based on Guided Attention and Scale Perception

Yufei Rao; Wei Guo; Xiaoyan Song; Gang Liang; Fengqing Cui; Binjun Ou

doi:10.3788/LOP242530

Laser & Optoelectronics Progress, Volume. 62, Issue 16, 1637003(2025)

Defect Detection Algorithm of Infrared Insulator Image Based on Guided Attention and Scale Perception

Yufei Rao^1、*, Wei Guo², Xiaoyan Song¹, Gang Liang¹, Fengqing Cui¹, and Binjun Ou³

Author Affiliations

¹Electric Power Science Research Institute, State Grid Henan Electric Power Company, Zhengzhou 450003, Henan , China

²State Grid Henan Electric Power Company, Zhengzhou 450000, Henan , China

³State Key Laboratory of New Energy Power System, North China Electric Power University, Beijing 102206, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(21)

Fig. 1. GASPNet structure

Download full size

Fig. 2. GAM structure

Download full size

Fig. 3. Network structures of spatial and channel attention in GAM

Download full size

Fig. 4. Structures of FPN and improved FEFN

Download full size

Fig. 5. Examples of partial defects in the dataset

Download full size

Fig. 6. Examples of the CPLID dataset

Download full size

Fig. 7. Comparison of P-R curves between original YOLOv11 and GASPNet

Download full size

Fig. 8. Comparison of detection results before and after adding GAM module in complex background noise interference scenarios

Download full size

Fig. 9. Comparison of detection results before and after adding FEFN in variable defect scale scenarios

Download full size

Fig. 10. Comparison of detection results before and after replacing EIoU loss function in scenarios with small defect scale

Download full size

Fig. 11. Comparative experimental results

Download full size

Fig. 12. Robustness verification visual experiment results

Download full size

Fig. 13. Visual analysis of activation heat maps of different attentional mechanisms by gradient weighting

Download full size

Table 1. Different input resolution comparison experiments
View table
Table 1. Different input resolution comparison experiments
Input resolution mAP /% Params /10⁶ FLOPs /10⁹ FPS /（frame/s）
640×640 94.8 2.9 8.2 95.3
1280×860 95.0 2.9 51.7 23.6

Table 2. Ablation results of self-made insulator defect detection dataset

View table

Table 2. Ablation results of self-made insulator defect detection dataset

GAM	FEFN	EIoU	AP /%			mAP /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）
GAM	FEFN	EIoU	broken	drop	flashover	mAP /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）
‒	‒	‒	94.3	93.8	92.4	93.5	2.6	6.5	99.4
√	‒	‒	95.0	94.3	93.0	94.1	2.7	6.9	97.5
‒	√	‒	94.8	94.4	92.5	93.9	2.9	7.6	96.9
‒	‒	√	94.7	93.8	92.9	93.8	2.6	6.5	99.4
√	√	√	95.6	94.7	94.1	94.8	2.9	8.2	95.3

Table 3. Comparative experiments of different loss functions
View table
Table 3. Comparative experiments of different loss functions
Loss function P R mAP
GIoU 94.1 92.8 93.7
DIoU 94.4 93.0 94.1
SIoU 93.5 92.5 93.3
CIoU 93.9 92.4 93.5
EIoU 95.2 94.1 94.8

Table 4. Comparative experimental results

View table

Table 4. Comparative experimental results

Detection algorithm	P /%	R /%	mAP /%	Params /10⁶	FLOPs /10⁹	FPS /（frame/s）
Faster-RCNN	87.3	86.4	86.8	28.3	168.4	6.1
SSD	63.7	62.9	63.3	23.8	72.9	70.9
YOLOv5	88.5	87.8	88.1	7.1	15.5	54.3
YOLOX	89.4	88.5	88.9	9.0	26.8	56.8
YOLOv8	90.2	89.3	89.6	11.2	28.6	78.3
YOLOv10	90.3	89.4	89.8	7.2	21.6	90.8
YOLOv11	93.9	92.4	93.5	2.6	6.5	99.4
GASPNet	95.2	94.1	94.8	2.9	8.2	95.3

Table 5. Comparison of experimental results of CPLID dataset

View table

Table 5. Comparison of experimental results of CPLID dataset

Detection algorithm	P	R	mAP
Faster R-CNN	81.2	78.5	79.4
CenterNet	90.5	87.1	88.9
YOLOv8	96.3	92.5	96.4
YOLOv11	98.9	97.3	98.2
YOLOv5s-KE^［27］			91.1
CDDCR-YOLOv8^［28］	96.9	94.3	97.5
ML-YOLOv5^［29］	96.5	94.7	97.0
PDDD-Net^［30］			83.3
YOLOX++S^［31］			96.6
BC-YOLO^［32］	96.8	94.6	96.7
GC-YOLO^［33］	94.7	90.9	94.9
YOLOD^［34］			98.0
GASPNet	99.7	98.9	99.3

Table 6. Comparison experiment results of missed and false detections
View table
Table 6. Comparison experiment results of missed and false detections
Detection algorithm Defect type Defect quantity N_FN N_FP
YOLOv11 broken 3158 205 158
drop 1958 157 108
flashover 2453 221 147
GASPNet broken 3158 158 95
drop 1958 127 78
flashover 2453 172 123

Table 7. Embedded device performance comparison

View table

Table 7. Embedded device performance comparison

Detection algorithm	P /%	R /%	mAP /%	FPS /（frame/s）
Faster-RCNN	87.3	86.4	86.8	3.0
SSD	63.7	62.9	63.3	26.1
YOLOv5	88.5	87.8	88.1	27.6
YOLOX	89.4	88.5	88.9	25.8
YOLOv8	90.2	89.3	89.6	33.3
YOLOv10	90.3	89.4	89.8	38.7
YOLOv11	93.9	92.4	93.5	49.2
GASPNet	95.2	94.1	94.8	47.1

Table 8. Comparison of performance of different attention mechanisms
View table
Table 8. Comparison of performance of different attention mechanisms
Attention mAP /% Params /10⁶ FLOPs /10⁹ FPS /（frame/s）
93.5 2.6 6.5 99.4
SE 93.6 2.6 6.6 98.3
SAM 93.7 2.7 7.0 96.3
CBAM 93.6 2.6 6.7 97.9
GAM 94.1 2.7 6.9 97.5

Tools

Get Citation

Copy Citation Text

Yufei Rao, Wei Guo, Xiaoyan Song, Gang Liang, Fengqing Cui, Binjun Ou. Defect Detection Algorithm of Infrared Insulator Image Based on Guided Attention and Scale Perception[J]. Laser & Optoelectronics Progress, 2025, 62(16): 1637003

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites