Author Affiliations
1School of Automation and Electronic Information, Xiangtan University, Xiangtan, Hunan 411105, China2Key Laboratory of Intelligent Computing & Information Processing of Ministry of Education, Xiangtan University, Xiangtan, Hunan 411105, China;show less
Fig. 1. Insulator detection based on the YOLOv3 network
Fig. 2. Structure of the Darknet-53
Fig. 3. Structure of the DenseNet
Fig. 4. Structure of the densely connected network
Fig. 5. Structure of the SPP module
Fig. 6. Structuref the DSM-Darknet
Fig. 7. Insulator data set. (a) GS; (b) PS; (c) PPin; (d) PPost; (e) PC; (f) R
Fig. 8. Results of the K-means clustering
Fig. 9. Training loss curve of the insulator data set
Fig. 10. PR curves of different insulators
Fig. 11. ROC of different insulators
Fig. 12. mAP of different algorithms
Fig. 13. Detection effect of the multi-scale dense network
Category | Total | Train | TrainVal | Test |
---|
GS | 5085 | 3051 | 762 | 1272 | PS | 8872 | 5323 | 1330 | 2219 | PPin | 2245 | 1347 | 336 | 562 | PPost | 5956 | 3573 | 893 | 1490 | PC | 6839 | 4103 | 1025 | 1711 | R | 4745 | 2847 | 711 | 1187 |
|
Table 1. Number of different categories in the data set
Algorithm | Anchor | Input size | Dense Net | SPP | Multi-scale | Loss | GS /% | PS /% | PPin /% | PPost /% | PC /% | R /% | mAP /% | FPS /frame |
---|
YOLOv3 | | | | | | | 88.6 | 89.1 | 73.5 | 74.2 | 89.1 | 74.2 | 81.5 | 40 | Improve1 | √ | | | | | | 89.2 | 90.6 | 75.3 | 76.5 | 90.3 | 76.6 | 83.1 | 41 | Improve2 | √ | √ | | | | | 89.6 | 91.1 | 76.9 | 78.1 | 91.1 | 78.1 | 84.2 | 40 | Improve3 | √ | √ | √ | | | | 91.8 | 92.8 | 82.9 | 81.8 | 93.9 | 81.5 | 87.5 | 34 | Improve4 | √ | √ | √ | √ | | | 93.1 | 94.2 | 84.8 | 85.7 | 95.0 | 85.7 | 89.8 | 33 | Improve5 | √ | √ | √ | √ | √ | | 94.3 | 95.7 | 86.4 | 87.3 | 97.3 | 88.2 | 91.5 | 31 | Ours | √ | √ | √ | √ | √ | √ | 96.1 | 96.6 | 88.5 | 89.8 | 98.1 | 90.9 | 93.4 | 28 |
|
Table 2. Different strategies influence the results of the algorithm
Algorithm | Backbone | Iterations | Precision /% | Recall /% | mAP /% | FPS /frame |
---|
Faster R-CNN | VGG16-Net | 60000 | 79.1 | 77.5 | 74.6 | 7 | SSD | VGG16-Net | 60000 | 82.0 | 83.2 | 78.9 | 26 | YOLOv3 | DarkNet53 | 60000 | 83.7 | 83.5 | 81.5 | 40 | RetinaNet | ResNet101 | 60000 | 84.8 | 84.3 | 82.4 | 20 | Ours | DSM-DarkNet | 60000 | 94.5 | 96.0 | 93.4 | 28 |
|
Table 3. Test results of different algorithms