Author Affiliations
1State Key Laboratory Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China2Micro Optics Electronic Machine System Education Ministry Key Laboratory, Tianjin University, Tianjin 300072, Chinashow less
Fig. 1. Structure of the Darknet-53
Fig. 2. DOTA data set. (a) Original image; (b) cropped image
Fig. 3. Enhancement of the DOTA data set. (a) Original image; (b) enhanced image
Fig. 4. Improved network prediction structure1
Fig. 5. Improved network prediction structure2
Fig. 6. Optimized network of the receptive field
Fig. 7. Detection effect of different networks. (a) YOLOv3; (b) structure1; (c) structure2
Fig. 8. Loss curve during training
Fig. 9. Recognition effect of different networks. (a) Structure2; (b) optimize the network of the receptive field
Fig. 10. Detection results of different networks under the COCO data set. (a) YOLOv3 network; (b) optimize the network of the receptive field
Backbone | Top-1/% | Top-5/% | FPS /frame |
---|
Darknet-19 | 74.1 | 91.8 | 171 | ResNet-101 | 77.1 | 93.7 | 53 | ResNet-152 | 77.6 | 93.8 | 37 | Darknet-53 | 77.2 | 93.8 | 78 |
|
Table 1. Performance of different backbone networks
Network | Plane | Large-vehicle | Small-vehicle | Average |
---|
YOLOv3 | 98.0 | 85.8 | 82.4 | 88.6 | Improved structure1 | 98.3 | 90.0 | 85.6 | 91.3 | Improved structure2 | 97.8 | 88.9 | 96.9 | 94.5 |
|
Table 2. Recall rates of different networks unit: %
Network | Plane | Large-vehicle | Small-vehicle | Average |
---|
YOLOv3 | 97.4 | 80.0 | 81.4 | 86.3 | Improved structure1 | 97.0 | 83.0 | 85.0 | 88.3 | Improved structure2 | 96.5 | 81.1 | 87.1 | 88.2 |
|
Table 3. Precision rates of different networks unit: %
Network | Plane | Large-vehicle | Small-vehicle | Average |
---|
YOLOv3 | 64.3 | 27.5 | 11.6 | 34.5 | R-FCN | 72.9 | 31.9 | 14.2 | 39.7 | Improved structure2 | 89.6 | 57.2 | 61.0 | 69.3 | Receptive field optimization | 93.1 | 76.8 | 72.7 | 80.9 |
|
Table 4. Multi-category recall rates of different networks unit: %
Network | Plane | Large-vehicle | Small-vehicle | Average |
---|
YOLOv3 | 62.0 | 16.7 | 3.5 | 27.4 | R-FCN | 68.6 | 29.1 | 13.5 | 34.9 | Improved structure2 | 87.9 | 45.8 | 44.8 | 59.5 | Receptive field optimization | 90.3 | 37.7 | 49.4 | 59.0 |
|
Table 5. Multi-class precision rates of different networks unit: %
Network | Volume /Mb | Time consuming /s |
---|
YOLOv3 | 246.3 | 0.063 | R-FCN | 102.5 | 0.180 | Improved structure2 | 242.7 | 0.085 | Receptive field optimization | 239.4 | 0.083 |
|
Table 6. Basic parameters of different networks
Network | Small | Medium | Large | Average |
---|
YOLOv3 | 24.0 | 48.2 | 61.1 | 44.4 | Receptive field optimization | 36.2 | 58.2 | 65.5 | 53.3 |
|
Table 7. Recall rates of different networks under the COCO data set unit: %
Network | Small | Medium | Large | Average |
---|
YOLOv3 | 14.2 | 34.1 | 46.4 | 31.6 | Receptive field optimization | 25.2 | 41.5 | 48.5 | 38.4 |
|
Table 8. Precision rates of different networks under the COCO data set unit: %