Acta Photonica Sinica, Volume. 53, Issue 8, 0810001(2024)
A Multi-scale Hierarchical Residual Network-based Method for Tiny Object Detection in Optical Remote Sensing Images
Figures & Tables(15)
Fig. 3. Diagram of calculation factors for bounding box regression indicators
Fig. 4. Partial image instances and tiny objects in the AI-TODv2 dataset
Fig. 5. The generated heatmap at the P2 layer based on the CSPNet module and MHRM module
Fig. 6. Visual examples of detection results of different detection algorithms on AI-TODv2 dataset
Fig. 7. Visual examples of detection results of different detection algorithms on the TinyPerson dataset
Table 1. Comparison of the combination results of MHRM module and CSPNet module with P3-5 layer and P2-4 layer
View tableView in ArticleTable 1. Comparison of the combination results of MHRM module and CSPNet module with P3-5 layer and P2-4 layer
Backbone module Layers AP50 AP75 mAP CSPNet （baseline）［ 21 ］P3， P4， P5 62.2 21.3 28.0 CSPNet （baseline） P2， P3， P4 67.6 25.2 31.4 MHRM （ours） P3， P4， P5 63.9 25.2 30.2 MHRM （ours） P2， P3， P4 67.6 27.6 32.5
Table 2. Comparison between localization loss functions
View tableView in ArticleTable 2. Comparison between localization loss functions
Method AP50 AP75 mAP CIoU［ 26 ］ （baseline）67.6 27.6 32.5 IoU loss + CD， β=0.5 （ours） 66.1 26.1 31.5 IoU loss + CD， β=1.0 （ours） 67.3 27.8 32.6 IoU loss + CD + BD， β=0.5 （ours） 66.8 29.1 33.1 IoU loss + CD + BD， β=1.0 （ours） 68.1 29.2 33.5
Table 3. Comparison of three combinations of values for the number of bottleneck block stacks n in MHRM at layers C2， C3， C4 and C5
View tableView in ArticleTable 3. Comparison of three combinations of values for the number of bottleneck block stacks n in MHRM at layers C2， C3， C4 and C5
n（C2， C3， C4， C5） Params （M） GFLOPs （G） AP50 AP75 mAP Inference/ms 1， 2， 2， 1 28.63 181.1 67.2 29.1 33.1 14.5 1， 3， 3， 1 30.93 199.4 67.4 28.9 33.1 15.8 1， 2， 3， 1 30.47 190.5 68.1 29.2 33.5 15.3
Table 4. Comparison of five different backbone modules
View tableView in ArticleTable 4. Comparison of five different backbone modules
Backbone Params （M） GFLOPs （G） AP50 AP75 mAP Inference/ms ResNet［ 5 ］38.30 191.6 65.6 23.1 29.6 17.1 CSPNet［ 21 ］34.79 183.8 67.6 25.2 31.4 16.8 SwinT-t*［ 28 ］35.60 671.9 63.1 21.8 28.2 23.9 Res2Net［ 24 ］35.66 181.6 65.0 25.0 30.6 17.1 MHRM （ours） 30.47 190.5 67.6 27.6 32.5 15.3
Table 5. Comparison of detection results between MHRM-YOLO ported to the YOLOv8 detection algorithm and the benchmark algorithm
View tableView in ArticleTable 5. Comparison of detection results between MHRM-YOLO ported to the YOLOv8 detection algorithm and the benchmark algorithm
Method Layers Params （M） GFLOPs （G） AP50 AP75 mAP Inference/ms YOLOv8［ 29 ］P3， P4， P5 43.6 257.5 60.3 24.0 29.0 15.8 YOLOv8［ 29 ］P2， P3， P4 34.7 309.8 64.0 25.7 30.7 18.8 MHRM-YOLO （ours） P3， P4， P5 50.1 225.9 63.0 27.2 31.4 10.1 MHRM-YOLO （ours） P2， P3， P4 33.8 248.9 65.7 29.4 33.3 12.1
Table 6. Comparison of detection performance between benchmark algorithms and MHRM on the AI-TODv2 test set
View tableView in ArticleTable 6. Comparison of detection performance between benchmark algorithms and MHRM on the AI-TODv2 test set
Method Backbone Params （M） GFLOPs AP50 AP75 mAP Inference/ms CasCade-RCNN/FPN［ 30 ］ResNet50［ 5 ］69.17 324.6 26.1 9.1 11.8 44.4 Faster-RCNN/FPN［ 31 ］ResNet50 41.38 269.1 24.3 9.2 11.2 39.4 Sparce RCNN［ 32 ］ResNet50 100.60 192.2 24.5 6.2 9.7 49.8 DetectoRS［ 33 ］ResNet50 124.00 236.3 29.3 12.7 14.8 136.1 Mask RCNN［ 34 ］SwinT-s［ 29 ］47.43 297.6 26.1 9.1 12.1 78.1 YOLOX［ 35 ］CSPNet［ 21 ］54.13 243.3 53.7 15.3 20.6 47.4 YOLOv5［ 19 ］CSPNet 46.10 168.6 62.2 21.3 28.0 11.7 YOLOv6［ 36 ］CSPBep 59.54 235.2 46.4 17.4 21.6 15.0 YOLOv7［ 23 ］ELAN 36.50 161.4 42.3 19.5 18.0 9.7 YOLOv8［ 29 ］C2f 43.60 257.5 60.3 24.0 29.0 15.8 MHRM-YOLO （ours） MHRM 30.47 190.5 68.1 29.2 33.5 15.3
Table 7. The AP and mAP performance of different algorithms on different categories on the AI-TODv2 test dataset
View tableView in ArticleTable 7. The AP and mAP performance of different algorithms on different categories on the AI-TODv2 test dataset
Method AI BR ST SH SP VE PE WM AP50 mAP CasCade-RCNN/FPN［ 30 ］21.2 6.2 19.7 21.9 8.2 12.6 5.0 0.0 26.1 11.8 Faster-RCNN/FPN［ 31 ］21.3 7.1 17.9 18.9 8.7 11.0 4.6 0.0 24.3 11.2 Sparce RCNN［ 32 ］9.1 9.5 13.2 26.1 5.5 7.2 3.6 3.5 24.5 9.7 DetectoRS［ 33 ］26.6 14.0 18.8 24.6 14.4 13.3 6.2 0.01 29.3 14.8 Mask RCNN［ 34 ］22.7 8.9 18.4 22.4 11.0 10.2 2.8 1.3 26.1 12.1 Cascade R-CNN/GPM ［ 18 ］24.3 23.9 42.5 46.0 30.1 30.8 9.4 9.4 58.6 27.1 FSANet-SwinT*［ 11 ］30.9 15.1 35.0 40.3 19.8 24.9 8.9 5.6 52.8 22.6 YOLOX［ 35 ］30.9 14.3 39.7 35.4 11.0 29.9 12.0 5.4 53.7 20.6 YOLOv5［ 19 ］38.6 18.5 41.8 43.8 24.3 33.1 15.9 7.7 62.2 28.0 YOLOv6［ 36 ］37.4 15.9 36.6 33.9 10.0 24.1 12.3 3.0 46.4 21.6 YOLOv7［ 23 ］25.7 9.0 35.9 31.0 23.5 28.1 10.6 1.2 42.3 18.0 YOLOv8［ 29 ］42.1 22.6 43.6 46.4 19.5 34.1 15.5 8.4 60.3 29.0 MHRM-YOLO （ours） 43.6 23.6 48.7 51.0 30.6 39.2 20.4 10.5 68.1 33.5
Table 8. Comparison of detection algorithms on the TinyPerson dataset
View tableView in ArticleTable 8. Comparison of detection algorithms on the TinyPerson dataset
Method Params （M） GFLOPs AP50 AP75 mAP YOLOv5［ 19 ］46.11 431.2 42.5 6.1 13.7 YOLOv6［ 36 ］59.50 600.2 33.3 4.2 10.9 YOLOv7［ 23 ］36.50 412.8 41.0 8.6 12.7 YOLOv8［ 29 ］43.60 663.0 42.1 6.5 14.5 MHRM-YOLO （ours） 30.45 486.8 46.2 6.4 15.5
Tools
Get Citation
Copy Citation Text
Xiangjin ZENG, Genghuan LIU, Jianming CHEN, Jiazhen DOU, Zhenbo REN, Jianglei DI, Yuwen QIN. A Multi-scale Hierarchical Residual Network-based Method for Tiny Object Detection in Optical Remote Sensing Images[J]. Acta Photonica Sinica, 2024, 53(8): 0810001
Paper Information
Category:
Received: Jan. 11, 2024
Accepted: Mar. 5, 2024
Published Online: Oct. 15, 2024
The Author Email: DI Jianglei (jiangleidi@gdut.edu.cn), QIN Yuwen (qinyw@gdut.edu.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20