Optics and Precision Engineering, Volume. 31, Issue 16, 2465(2023)
Remote sensing multi-scale object detection based on multivariate feature extraction and characterization optimization
Figures & Tables(20)
Fig. 1. Network structure of multivariate feature extraction and characterization optimization
Fig. 8. Experiential results of our proposed method（shown on the down） and Baseline（shown on the up） on DIOR dataset
Fig. 10. Experiential results of our proposed method（shown on the down） and Baseline（shown on the up） on HRRSD dataset
Fig. 12. Experiential results of our proposed method（shown on the down） and Baseline（shown on the up） on RSOD dataset
Table 1. Performance comparison of different algorithms on DIOR dataset
View tableView in ArticleTable 1. Performance comparison of different algorithms on DIOR dataset
Faster RCNN［ 1 ］Mask
R-CNN［
2 ］YOLOx［ 3 ］Retina-Net［ 4 ］PANet［ 31 ］CSFF［ 7 ］CBD-E［ 32 ］Corner-Net［ 14 ］Baseline 本文方法
MFC
D7 60.0 60.4 53.1 62.4 61.4 62.7 63.5 64.3* 51.4 54.5 D8 75.6 76.3 55.6 78.6 72.1 82.6* 76.2 81.6 62.3 78.1 D9 62.3 62.5 57.2 62.8 66.7 73.2 65.3 76.3* 60.0 67.8 D10 76.0 76.0 77.5 78.6 72.0 78.2 79.3 79.5 75.3 80.5 *D11 76.8 75.9 73.3 76.6 73.4 81.6 *79.5 79.5 74.0 76.6 D12 46.4 46.5 48.2 49.9 45.3 50.7 47.5 26.1 51.3 54.3 *D13 57.2 57.4 53.7 59.6 56.9 59.5 59.3 60.6 55.8 61.5 *D14 71.8 71.8 83.9 71.1 71.7 73.3 69.1 37.6 84.7 87.4 *D15 68.3 68.3 62.2 68.4 70.4 63.4 69.7 70.7 *56.2 65.3 D16 53.8 53.7 61.4 45.8 62.0 58.5 64.3 45.2 68.9 72.6 *D17 81.1 81.0 86.3 81.3 80.9 85.9 84.5 84.0 86.3 88.2 *D18 59.5 62.3 61.4 55.2 57.0 61.9 59.4 57.1 53.4 62.4 *D19 43.1 43.0 47.3 44.4 47.2 42.9 44.7 43.0 45.1 47.9 *D20 81.2 81.0 84.5 85.5 84.5 86.9 83.1 75.9 75.7 87.2* mAP/%↑ 65.1 65.2 65.9 66.1 66.1 68.0 67.8 64.9 66.2 70.9* FPS↑ 11.2 12.9 33.2 16.4 14.5 - - 26.7 30.5 28.1 FLOPs↓ 38.7 G 38.4 G 11.4 G 25.2 G 26.9 G - - 12.8 G 12.5 G 12.7 G
Table 1. Performance comparison of different algorithms on DIOR dataset
View tableView in ArticleTable 1. Performance comparison of different algorithms on DIOR dataset
Faster RCNN［ 1 ］Mask
R-CNN［
2 ］YOLOx［ 3 ］Retina-Net［ 4 ］PANet［ 31 ］CSFF［ 7 ］CBD-E［ 32 ］Corner-Net［ 14 ］Baseline 本文方法
MFC
D1 54.0 53.9 72.5 53.3 60.2 57.2 54.2 58.8 78.1 87.4* D2 74.5 76.6 76.3 77.0 72.0 79.6 77.0 84.2* 71.2 83.6 D3 63.3 63.2 74.1 69.3 70.6 70.1 71.5 72.0 72.6 74.3* D4 80.7 80.9 77.4 85.0 80.5 87.4 87.1 80.8 87.1 87.8* D5 44.8 40.2 40.8 44.1 43.6 46.1 44.6 46.4 38.8 47.2* D6 72.5 72.5 70.5 73.2 72.3 76.6 75.4 75.3 75.3 77.3*
Table 2. Performance comparison of different algorithms on HRRSD dataset
View tableView in ArticleTable 2. Performance comparison of different algorithms on HRRSD dataset
Faster RCNN［ 1 ］GACL
FasterRCNN［
29 ］YOLOx［ 3 ］RetinaNet［ 4 ］PANet［ 31 ］CornerNet［ 14 ］Baseline 本文方法
MFC
H1 90.5 90.8 91.3 92.6 93.7 88.5 96.6 98.4* H2 86.1 88.5 85.6 86.7 87.2 89.2 90.7 92.2* H3 87.3 89.2 86.7 89.1 90.5 93.3 94.1 95.0* H4 86.7 87.2 86.3 87.5 89.3 89.5 88.4 91.1* H5 80.7 80.8 85.3 84.9 87.7 90.6 91.8 93.8* H6 48.2 49.7 71.7 51.3 58.2 72.6 71.4 75.7* H7 89.9 90.7 87.4 89.6 92.3 94.0 94.2 96.9* H8 90.1 89.7 90.2 89.9 89.7 89.9 91.5 94.8* H9 86.2 85.6 85.1 86.2 88.4 88.5 88.7 92.3* H10 86.5 86.9 85.4 87.7 88.9 89.2 90.3 93.8* H11 89.4 88.2 87.5 87.4 89.4 90.8 91.9 94.9* H12 74.7 75.0 74.1 75.3 79.3 79.4 80.0 85.8* H13 66.2 65.3 65.3 68.3* 66.5 66.8 63.2 67.6 mAP/%↑ 81.7 82.1 83.2 82.8 84.7 86.3 87.1 90.2* FPS↑ 14.5 - 48.9 23.2 20.9 42.2 45.5 43.7 FLOPs↓ 38.7 G - 11.4 G 25.2 G 26.9 G 12.8 G 12.5 G 12.7 G
Table 3. Performance comparison of different algorithms on RSOD dataset
View tableView in ArticleTable 3. Performance comparison of different algorithms on RSOD dataset
Faster RCNN［ 1 ］Mask
R-CNN［
2 ］YOLOx［ 3 ］RetinaNet［ 4 ］CBD-E［ 32 ］CornerNet［ 14 ］Baseline 本文方法MFC R1 79.6 81.2 87.5 82.2 95.8* 92.1 91.4 94.4 R2 94.7 93.4 95.3 99.7 99.7 96.8 96.3 99.7* R3 85.0 85.9 83.6 93.8 88.8 90.0 90.1 94.1* R4 93.8 93.5 91.2 97.1 94.2 93.5 94.9 99.5* mAP/%↑ 88.3 88.5 89.4 93.2 94.2 93.1 93.2 96.9* FPS↑ 15.9 17.5 61.3 27.6 - 56.3 61.1 58.8 FLOPs↓ 38.7 38.4G 11.4G 25.2G - 12.8G 12.5G 12.7G
Table 4. Ablation experiments on RSOD datasets（MFE）
View tableView in ArticleTable 4. Ablation experiments on RSOD datasets（MFE）
Method mAP/%↑ FPS↑ FLOPs↓ Baseline 93.2 61.5 12.48 G +跨组连接 94.9 59.9 12.64 G +跨组连接+SE 95.2 59.7 12.69 G +跨组连接+SE+CM 95.7 59.4 12.71 G
Table 5. Ablation experiments on RSOD datasets（COS）
View tableView in ArticleTable 5. Ablation experiments on RSOD datasets（COS）
Method mAP/%↑ FPS↑ FLOPs↓ Baseline 93.2 61.5 12.48 G +标签优化 93.5 61.4 12.49 G +标签优化+CPA 94.1 60.1 12.52 G
Table 6. Contrast experiments on RSOD datasets（SE、CA）
View tableView in ArticleTable 6. Contrast experiments on RSOD datasets（SE、CA）
Method mAP/%↑ FPS↑ FLOPs↓ Baseline 93.2 61.5 12.48 G +MFE（with SE） 95.7 59.4 12.71 G +MFE（with CBAM） 95.8 59.4 12.72 G +COS（with CA［ 26 ］）94.1 60.1 12.52 G +COS（with CA［ 27 ］）93.9 60.2 12.50 G
Table 7. Ablation experiments on DIOR、HRRSD and RSOD datasets
View tableView in ArticleTable 7. Ablation experiments on DIOR、HRRSD and RSOD datasets
消融实验 DIOR HRRSD RSOD FLOPs MFE COS mAP/% FPS MFE COS mAP/% FPS MFE COS mAP/% FPS Baseline - - 66.2 30.5 - - 87.1 45.5 - - 93.2 61.5 12.5 G √ √ - 69.3 29.2 √ - 89.4 43.1 √ - 95.7 59.4 12.7 G √ √ √ 70.9 28.1 √ √ 90.2 43.7 √ √ 96.9 58.8 12.7 G
Tools
Get Citation
Copy Citation Text
Yuebo MENG, Fei WANG, Guanghui LIU, Shengjun XU. Remote sensing multi-scale object detection based on multivariate feature extraction and characterization optimization[J]. Optics and Precision Engineering, 2023, 31(16): 2465
Paper Information
Category: Information Sciences
Received: Nov. 10, 2022
Accepted: --
Published Online: Sep. 5, 2023
The Author Email: LIU Guanghui (guanghuil@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20