Laser & Optoelectronics Progress, Volume. 62, Issue 4, 0428002(2025)
Strip Object Detection Method for Multiscale Optical Remote Sensing Images Without Anchors
Figures & Tables(14)
Fig. 7. Loss variation curves during training reasoning process of different methods
Fig. 9. Comparison of detection results between proposed method and other methods on DIOR dataset. (a) Faster R-CNN method; (b) CenterNet method; (c) SSD method; (d) RetinaNet method; (e) proposed method
Table 1. Parameters setting
View tableTable 1. Parameters setting
Parameter Value Parameter Value Input image size 800×800 Mixed precision training True Freeze train True Pretrained True Freeze batchsize 16 Freeze epoch 50 Unfreeze batchsize 8 Unfreeze epoch 100 Init_lr 0.0005 Lr_decay_type Cos Optimizer_type Adam Momentum 0.9
Table 2. Effectiveness experiments of proposed CSASPP module at different dilation rates
View tableTable 2. Effectiveness experiments of proposed CSASPP module at different dilation rates
Model Backbone Dilation rate IOU is 0.5 IOU is 0.75 mAP /% MR-2 /% mAP /% MR-2 /% ASPP+CA+SP-1 Resnet50 （1，3，6） 79.08 33.10 55.13 55.10 ASPP+CA+SP-2 Resnet50 （3，6，9） 79.90 32.35 55.48 54.85 ASPP+CA+SP-3 Resnet50 （6，12，18） 79.19 32.80 54.32 55.30 ASPP+CA+SP-4 Resnet50 （9，18，36） 79.71 32.35 55.28 54.70
Table 3. Ablation experiment
View tableTable 3. Ablation experiment
Model Backbone Dilation rate IOU is 0.5 IOU is 0.75 mAP /% MR-2 /% mAP /% MR-2 /% ASPP［31］ Resnet50 （6，12，18） 77.98 34.20 52.77 56.95 ASPP+ Resnet50 （3，6，9） 79.55 32.80 53.88 55.95 ASPP+CA Resnet50 （3，6，9） 75.31 39.45 46.56 63.40 ASPP+SP Resnet50 （3，6，9） 79.34 32.80 54.53 55.90 ASPP+CA+SP Resnet50 （3，6，9） 79.90 32.35 55.48 54.85
Table 4. Performance of the proposed method with other methods at different thresholds
View tableTable 4. Performance of the proposed method with other methods at different thresholds
Method Backbone IOU is 0.5 IOU is 0.75 Number of parameters /106 GFLOPS /109 mAP /% MR-2 /% mAP /% MR-2 /% Faster R-CNN［4］ Resnet50 64.66 47.40 38.09 66.80 137.078 547.392 CenterNet［11］ Resnet50 71.27 43.80 44.03 65.10 32.665 171.428 SSD［32］ Resnet50 75.96 41.60 52.34 61.20 14.211 104.309 RetinaNet［33］ Resnet50 78.06 32.50 60.68 50.65 36.724 266.322 Proposed method Resnet50 79.90 32.35 55.48 54.85 186.327 355.464
Table 5. AP values of different methods and categories (IOU is 0.5)
View tableTable 5. AP values of different methods and categories (IOU is 0.5)
Categories Faster R-CNN CenterNet SSD RetinaNet Proposed method mAP 64.66 71.27 75.96 78.06 79.90 Airplane 64.78 93.23 93.36 88.69 93.07 Vehicle 16.76 60.87 55.29 39.37 62.90 Ship 20.24 83.30 94.75 77.03 81.72 Storage tank 21.14 60.92 68.86 45.29 58.60 Wind mill 53.01 83.86 90.43 86.05 87.64 Chimney 86.62 87.41 92.71 94.84 89.17 Dam 70.65 50.10 62.43 84.51 70.79 Overpass 58.92 55.55 63.54 65.63 64.76 Bridge 35.73 41.65 39.33 44.66 55.08 Tennis court 87.45 93.89 95.47 95.15 94.16 Airport 83.17 56.89 62.69 81.72 86.83 Golf field 90.22 64.72 69.53 91.01 88.93 Ground track field 77.94 81.06 82.19 90.43 88.88 Stadium 93.40 85.95 89.30 97.63 95.40 Expressway service area 82.51 74.36 86.42 94.87 83.92 Expressway toll station 59.20 78.45 75.52 78.07 78.75 Harbor 61.39 50.36 66.85 67.49 58.20 Train station 59.83 43.26 48.22 56.61 78.63 Baseball field 87.11 91.56 92.05 91.64 92.18 Basketball court 83.05 88.03 90.20 90.49 88.45
Tools
Get Citation
Copy Citation Text
Qi He, Hao Shen. Strip Object Detection Method for Multiscale Optical Remote Sensing Images Without Anchors[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0428002
Paper Information
Category: Remote Sensing and Sensors
Received: May. 9, 2024
Accepted: Jun. 27, 2024
Published Online: Feb. 18, 2025
The Author Email:
CSTR:32186.14.LOP241242
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20