Author Affiliations
1 Institute of Geospatial Information, Information Engineering University, Zhengzhou, Henan 450000, China2 78123 Troops, Chengdu, Sichuan 610000, Chinashow less
Fig. 1. Framework of SSD algorithm
Fig. 2. Framework of improved SSD algorithm
Fig. 3. Comparison of feature maps before and after integration. (a) Input image; (b) output of dense block2; (c) output of dense block3; (d) output of dense block2 with feature integration; (e) output of dense block3 with feature integration; (f) output of dense block4
Fig. 4. Interface of training sample online acquisition system. (a) Superimposed main airport point data; (b) aircraft sample collection
Fig. 5. Size of each target in sample set
Fig. 6. Decay curve of learning rate
Fig. 7. Comparison of total loss and precision between transfer training and random initialization. (a) Total loss varies with number of iterations; (b) MAPRIoU=0.50 varies with number of iterations
Fig. 8. Comparison of precisions of improved SSD algorithm and other algorithms varying with number of iterations. (a) MAP; (b) MAPlarge; (c) MAPmedium; (d) MAPsmall; (e) MAPRIoU=0.50; (f) MAPRIoU=0.75
Fig. 9. Comparison of improved SSD algorithm and other algorithms in detection effect. (a) Faster R-CNN+ResNet101; (b) R-FCN+ResNet101; (c) improved SSD algorithm
Metric | Remarks |
---|
MAP | MAP at RIoU in {0.5+0.05×m,m=0,1,…,9} (primary challenge metric) | | MAP at RIoU=0.50 (pascal VOC metric) | | MAP at RIoU=0.75 (strict metric) | MAPsmall | MAP for small targets: Sarea<(32 pixel)2 | MAPmedium | MAP for medium targets: (32 pixel)2≤Sarea≤(96 pixel)2 | MAPlarge | MAP for large targets: Sarea>(96 pixel)2 |
|
Table 1. Main metrics
Data set | Class | Target amount | Percentage /% |
---|
Small | Medium | Large | Total | | Small | Medium | Large |
---|
Training set | airplane | 1204 | 1542 | 431 | 4401 | 27.36 | 35.04 | 9.79 | | playground | 178 | 516 | 530 | 4.04 | 11.72 | 12.04 | Validation set | airplane | 390 | 427 | 74 | 1169 | 33.36 | 36.53 | 6.33 | | playground | 27 | 120 | 131 | 2.31 | 10.27 | 14.21 | Test set | airplane | 940 | 366 | 111 | 1892 | 49.68 | 19.34 | 5.87 | | playground | 133 | 294 | 48 | 7.03 | 15.54 | 2.54 |
|
Table 2. Sample set statistics
Method | Parameter /MB | Time overhead /ms | Metric /% |
---|
MAP | MAPlarge | MAPmedium | MAPsmall | | |
---|
SSD+Inceptionv2 | 53.4 | 24.8 | 47.15 | 69.48 | 50.16 | 11.56 | 85.08 | 48.95 | Faster R-CNN+ResNet50 | 173.3 | 108.6 | 47.12 | 72.81 | 48.91 | 10.16 | 82.72 | 50.59 | Faster R-CNN+ResNet101 | 249.5 | 117.5 | 50.50 | 73.06 | 53.51 | 13.76 | 85.84 | 55.03 | R-FCN+ResNet101 | 258.2 | 79.3 | 51.17 | 74.69 | 51.43 | 16.01 | 87.20 | 55.86 | Improved SSD algorithm | 59.8 | 71.8 | 54.14 | 73.31 | 54.32 | 21.16 | 90.55 | 57.38 |
|
Table 3. Comparison of calculation time and precision on validation set
Methods | Metric /% |
---|
MAP | MAPlarge | MAPmedium | MAPsmall | | |
---|
SSD+Inceptionv2 | 35.85 | 62.73 | 43.55 | 19.52 | 77.72 | 24.07 | Faster R-CNN+ResNet50 | 28.72 | 61.53 | 34.67 | 13.43 | 68.87 | 17.60 | Faster R-CNN+ResNet101 | 36.05 | 61.55 | 42.83 | 21.74 | 76.83 | 27.69 | R-FCN+ResNet101 | 36.70 | 59.18 | 43.96 | 22.91 | 77.30 | 28.23 | Improved SSD algorithm | 45.18 | 65.31 | 50.68 | 31.65 | 83.95 | 42.15 |
|
Table 4. Comparison of precision on test set