Fig. 1. Faster R-CNN model for detecting images of insulation piercing connectors and bolts
Fig. 2. RPN model
Fig. 3. Anchors of different scales and lengths
Fig. 4. RoI pooling network
Fig. 5. Model structure diagram. (a) Faster R-CNN model; (b) improved Faster R-CNN model
Fig. 6. Residual block structure diagram of ResNet
Fig. 7. Detection results of improved Faster R-CNN model for connectors and bolts under different conditions. (a) Vertically inward bolt; (b) downward bolt; (c) upward bolt; (d) bolt shielded by its own wire clip
Fig. 8. Detection network structure of insulation piercing connectors and bolts on the transmission line based on improved Faster R-CNN
Fig. 9. Detection flow chart of insulation piercing connectors and bolts on the transmission line based on improved Faster R-CNN
Number ofimages | AP of insulationpiercing connector /% | AP ofbolt/% | mAP /% |
---|
1500 | 91.8 | 90.2 | 91.0 | 3000 | 93.3 | 91.5 | 92.4 |
|
Table 1. Comparison of results obtained from different training sample amount
Model | AP of insulationpiercingconnector /% | AP ofbolt /% | mAP /% |
---|
Faster R-CNN+ZFNet | 88.2 | 86.4 | 87.3 | Faster R-CNN+VGG-16 | 90.3 | 88.9 | 89.6 | Faster R-CNN+ResNet50 | 93.3 | 91.5 | 92.4 |
|
Table 2. Comparison of Faster R-CNN model results based on three different networks
Model | Number ofproposals | Batch size inRPN stage | Batch size in 2nd stage | mAP /% |
---|
| 300 | 128 | 64 | 92.4 | Faster R-CNN+ResNet50 | 250 | 128 | 64 | 91.1 | | 200 | 128 | 64 | 90.3 | | 150 | 128 | 64 | 88.5 |
|
Table 3. Comparison of effects of different number of proposals after first stage NMS on mAP
Model | Number ofproposals | Batch sizein RPN stage | Batch size in 2nd stage | mAP /% |
---|
| 300 | 128 | 64 | 92.4 | Faster R-CNN+ResNet50 | 300 | 64 | 32 | 91.6 | | 300 | 32 | 16 | 89.4 |
|
Table 4. Comparison of effects of different batch sizes on mAP
Model | AP of insulationpiercing connector /% | AP of bolt/% | mAP /% | Time per image /ms |
---|
Faster R-CNN+ResNet50 | 93.3 | 91.5 | 92.4 | 2639 | SSD | 84.5 | 80.3 | 82.3 | 68 | YOLO v3 | 85.6 | 83.5 | 84.6 | 44 |
|
Table 5. Comparison of effects of different object detection models on mAP