Optics and Precision Engineering, Volume. 33, Issue 10, 1672(2025)
Ultra lightweight SAR image small object detection network
Figures & Tables(18)
Fig. 1. Ultra lightweight framework for small object detection in SAR images
Fig. 8. Object detection effect of the proposed method in large scale scenarios
Fig. 9. Visual comparison of object detection results using various methods on multiple ship datasets
Table 1. Object size distribution in the dataset
View tableView in ArticleTable 1. Object size distribution in the dataset
数据集 目标总数 小尺寸 中尺寸 大尺寸 目标数量 占比/% 目标数量 占比/% 目标数量 占比/% MSAR 60 398 56 562 93.65 3 066 5.08 770 1.27 SAR-Ship 50 885 48 831 95.97 2 002 3.93 52 0.10 AIR-SARShip-2.0 2 040 2 016 98.82 23 1.13 1 0.05 SSDD 322 251 77.95 68 21.12 3 0.93 HRSID 16 950 16 858 99.46 83 0.49 9 0.05
Table 2. Detection performance of DESD-Head on SSDD dataset
View tableView in ArticleTable 2. Detection performance of DESD-Head on SSDD dataset
网络模型 DESD-Head FLOPs/G Params/M FPS mAP/% APS/% APM/% APL/% YOLOv8-n × 8.089 3.007 265.4 98.3 63.9 75.3 73.2 YOLOv8-n √ 6.525（↓19.33%） 2.362（↓21.45%） 274.2（↑3.21%） 98.0 64.3 73.5 74.4 YOLOv10-N × 6.532 2.266 178.9 97.6 61.7 73.2 64.7 YOLOv10-N √ 6.300（↓3.55%） 1.947（↓14.08%） 188.5（↑5.09%） 97.4 65.8 73.8 71.0
Table 3. Lightweight effect under different pruning ratios on the MSAR dataset
View tableView in ArticleTable 3. Lightweight effect under different pruning ratios on the MSAR dataset
网络模型 剪枝比例 FLOPs/G Params/M mAP/% APS/% APM/% APL/% 基线 - 8.089 3.007 84.6 70.6 80.1 69.1 M+D 0% 5.291 1.944 86.6 75.0 83.2 71.8 M+D 10% 4.724 1.041 85.8 74.1 82.3 63.5 M+D 20% 4.186 0.888 83.2 72.5 81.0 68.3 M+D 30% 3.661 0.742 78.9 69.5 78.4 58.7 M+D 40% 3.149 0.596 80.6 66.2 77.0 66.3
Table 4. Performance improvement results of student networks with different pruning ratios after distillation on the MSAR dataset
View tableView in ArticleTable 4. Performance improvement results of student networks with different pruning ratios after distillation on the MSAR dataset
学生网络 教师网络 FLOPs/G Params/M mAP/% APS/% APM/% APL/% 10%剪枝的M+D
（mAP 85.8）
YOLOv8-n 4.724 1.041 86.1（↑0.3） 75.1 82.7 73.1 YOLOv8-s 87.8（↑2.0） 75.1 83.7 75.9 YOLOv8-m 89.4 （↑3.6）76.1 85.1 78.3 YOLOv8-l 87.7（↑1.9） 74.7 84.0 76.2 YOLOv8-x 86.9（↑1.1） 74.5 83.3 73.4 20%剪枝的M+D
（mAP 83.2）
YOLOv8-n 4.186 0.888 86.9（↑3.7） 73.7 82.5 70.6 YOLOv8-s 89.0 （↑5.8）75.7 85.1 78.1 YOLOv8-m 85.7（↑2.5） 73.4 82.8 79.4 YOLOv8-l 84.1（↑1.9） 73.3 81.2 75.0 YOLOv8-x 83.8（↑0.6） 72.4 80.5 74.4 30%剪枝的M+D
（mAP 78.9）
YOLOv8-n 3.661 0.742 81.1 （↑2.9）65.9 76.7 81.1 YOLOv8-s 80.4（↑1.5） 67.0 77.9 80.4 YOLOv8-m 80.6（↑1.7） 66.2 77.9 80.6 YOLOv8-l 80.0（↑1.1） 65.7 76.1 79.6 YOLOv8-x 78.5（↓0.4） 62.8 72.0 74.4 40%剪枝的M+D
（mAP 80.6）
YOLOv8-n 3.149 0.596 80.4 （↓0.2）65.4 76.4 68.5 YOLOv8-s 80.3（↓0.3） 65.5 77.7 70.0 YOLOv8-m 76.9（↓3.7） 61.3 73.9 63.0 YOLOv8-l 76.5（↓4.1） 62.6 72.3 62.9 YOLOv8-x 75.7（↓4.9） 60.6 70.3 61.1
Table 5. Ablation experiments on the MSAR dataset
View tableView in ArticleTable 5. Ablation experiments on the MSAR dataset
网络
模型
MELA DESD-
Head
剪枝 蒸馏 FLOPs/G Params/M FPS/
（frame·s-1）
mAP/% APS/% APM/% APL/% Model-1 × × × × 8.089 3.007 263.5 84.6 70.6 80.1 69.1 Model-2 √ × × × 6.051 2.198 270.8 82.0 66.5 77.1 61.5 Model-3 × √ × × 6.525 2.362 230.2 89.6 78.0 86.2 74.8 Model-4 √ √ × × 5.291 1.944 231.9 86.6 75.0 83.2 71.8 Model-5 √ √ √ × 4.186 0.888 250.3 83.2 72.5 81.0 68.3 Model-6 √ √ √ √ 4.186 0.888 279.7 89.0 75.7 85.1 78.1
Table 6. Performance comparison results of different SAR images object detection methods on MSAR dataset
View tableView in ArticleTable 6. Performance comparison results of different SAR images object detection methods on MSAR dataset
网络模型 FLOPs/G Params/M FPS/（frame·s-1） mAP/% APS/% APM/% APL/% YOLOv5-n 4.130 1.761 125.0 80.8 73.9 74.4 80.7 YOLOv7 103.223 36.501 96.2 86.7 74.4 85.9 81.6 YOLOv7-tiny 13.022 6.007 109.9 68.9 62.9 72.3 61.0 YOLOv8-n 8.089 3.007 263.5 84.6 70.6 80.1 69.1 YOLOv9-T 7.599 1.971 42.5 77.6 71.2 80.5 77.8 YOLOv10-N 6.532 2.266 228.2 83.1 65.7 75.7 52.7 RT-DETR（r18） 57.290 18.860 36.7 76.8 51.4 71.1 55.4 本文方法 4.186 0.888 279.7 89.0 75.7 85.1 78.1
Table 7. Detection accuracy of various types of targets using different methods on the MSAR dataset
View tableView in ArticleTable 7. Detection accuracy of various types of targets using different methods on the MSAR dataset
网络模型 飞机 油罐 桥梁 船只 YOLOv5-n 93.8 83.8 74.6 71.0 YOLOv7 94.0 92.2 88.8 71.7 YOLOv7-tiny 88.4 85.9 46.9 54.4 YOLOv8-n 90.2 82.9 85.9 79.5 YOLOv9-T 88.4 87.5 67.9 66.5 YOLOv10-N 86.7 83.7 81.9 80.2 RT-DETR（r18） 96.5 79.6 77.3 54.0 本文方法 97.0 87.3 88.0 83.7
Table 8. Object detection results of different methods on SAR-Ship and AIR-SARShip-2.0 datasets
View tableView in ArticleTable 8. Object detection results of different methods on SAR-Ship and AIR-SARShip-2.0 datasets
网络模型 FLOPs/G Params/M SAR-Ship AIR-SARShip-2.0 FPS mAP/% FPS mAP/% YOLOv5-n 4.1 1.8 259.8 97.7 257.1 67.8 YOLOv7 103.2 36.5 80.2 94.7 83.0 63.0 YOLOv7-tiny 13.0 6.0 264.5 96.8 256.5 70.3 YOLOv8-n 8.1 3.0 259.9 97.7 263.3 75.7 YOLOv9-T 10.7 2.6 69.5 94.7 70.2 74.2 YOLOv10-N 6.5 2.3 168.8 93.1 201.7 64.5 RT-DETR（r18） 56.9 18.8 58.6 95.3 60.7 51.8 CenterNet++［ 39 ］89.4 20.3 52.3 94.9 48.4 73.9 GhostNet［ 32 ］8.7 3.6 204.1 91.7 199.9 49.0 Quant-Det［ 20 ］9.5 3.2 213.2 95.7 208.5 77.0 本文方法 4.2 0.9 281.3 98.1 283.5 82.5
Table 9. Object detection results of different methods on SSDD and HRSID datasets
View tableView in ArticleTable 9. Object detection results of different methods on SSDD and HRSID datasets
网络模型 FLOPs
/G
Params
/M
SSDD HRSID FPS mAP
/%
APS
/%
APM
/%
APL
/%
FPS mAP
/%
APS
/%
APM
/%
APL
/%
Cascade R-CNN［ 40 ］152.2 55.6 9.6 93.1 54.8 67.1 57.8 10.0 87.7 67.5 67.7 28.8 Mask Scoring R-CNN［ 41 ］206.7 48.1 13.4 94.7 55.6 65.3 51.2 12.6 87.6 65.3 65.8 22.2 Hybrid Task Cascade［ 42 ］278.5 69.3 6.3 94.5 55.9 68.4 59.2 7.8 87.7 69.0 71.2 38.1 RetinaNet［ 18 ］114.8 42.7 31.9 90.2 51.2 62.6 45.4 33.7 84.7 60.9 60.9 26.8 YOLOv5-n 4.1 1.8 254.2 97.2 62.1 74.6 69.1 256.6 90.3 50.0 73.5 65.7 YOLOv7 103.2 36.5 80.9 94.5 64.2 77.5 79.4 82.8 87.7 47.7 73.4 57.9 YOLOv7-tiny 13.0 6.0 263.6 96.2 62.7 78.4 72.0 262.3 87.0 45.3 71.5 57.2 YOLOv8-n 8.1 3.0 265.4 98.3 63.9 75.3 73.2 265.2 89.8 51.0 76.8 52.3 YOLOv9-T 10.7 2.6 72.6 97.0 63.3 75.7 72.8 79.1 88.2 52.7 73.1 62.8 YOLOv10-N 6.5 2.3 178.9 97.6 61.7 73.2 64.7 167.8 89.4 53.9 77.4 52.2 RT-DETR（r18） 56.9 18.8 60.6 97.2 64.6 76.4 77.6 60.0 91.3 51.4 75.7 60.5 本文方法 4.2 0.9 279.3 98.6 64.7 74.1 74.8 282.7 91.5 66.2 77.8 64.9
Tools
Get Citation
Copy Citation Text
Xiaomin YANG, Jun YANG. Ultra lightweight SAR image small object detection network[J]. Optics and Precision Engineering, 2025, 33(10): 1672
Paper Information
Category:
Received: Nov. 4, 2024
Accepted: --
Published Online: Jul. 23, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20