Oriented Object Detection in Remote Sensing Images Based on Feature Recombination

Youwei Wang; Ying Guo; Xiangying Shao; Jiyu Wang; Zhengwei Bao

doi:10.3788/AOS230957

Acta Optica Sinica, Volume. 44, Issue 6, 0628001(2024)

Oriented Object Detection in Remote Sensing Images Based on Feature Recombination

Youwei Wang^1,2, Ying Guo^1,2、*, Xiangying Shao^1,2, Jiyu Wang^1,2, and Zhengwei Bao^1,2

¹Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology, Nanjing University of Information Science & Technology, Nanjing 210044, Jiangsu, China

²School of Automation, Nanjing University of Information Science & Technology, Nanjing 210044, Jiangsu, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(18)

Fig. 1. Framework of Rotated RPN

Download full size

Fig. 2. Framework of PFR-Rotate

Download full size

Fig. 3. Framework of reshape model

Download full size

Fig. 4. Framework of feature recombination model

Download full size

Fig. 5. Midpoint offset definition method

Download full size

Fig. 6. Flow chart of ATT-ORPN

Download full size

Fig. 7. Structure diagram of double polarization attention head

Download full size

Fig. 8. Average precision (AP) for each category in Dior-R dataset for each experiment

Download full size

Fig. 9. Loss function and mAP comparison. (a) Loss function; (b) mAP

Download full size

Fig. 10. Comparison of visualization results for Dior-R dataset

Download full size

Fig. 11. Comparison of visualization results for HRSC2016 dataset

Download full size

Table 1. Comparison of FPN ablation experimental results

View table

Table 1. Comparison of FPN ablation experimental results

Baseline	SPP	CBAM	SE	Carafe	reshape	FR	mAP /%	mAP variation /%	Recall /%	f /（frame·s^-1）
√							59.54		—	—
√	√						62.26	↑2.72	71.07	56.6
√	√				√		62.32	↑2.78	71.11	54.3
√		√					62.42	↑2.88	71.29	56.7
√		√			√		62.49	↑2.95	71.31	51.9
√			√				62.77	↑3.23	70.18	56.6
√			√		√		62.86	↑3.32	70.35	55.3
√				√			62.90	↑3.36	70.42	54.6
√						√	63.03	↑3.49	70.68	51.3

Table 2. Comparison of polarization function ablation experimental results
View table
Table 2. Comparison of polarization function ablation experimental results
cls reg mAP /%
Eq .（2） Eq .（3） Eq .（4） Eq .（5） Eq .（6） Eq .（7）
√ √ 61.38
√ √ 60.87
√ √ 61.03
√ √ 61.97
√ √ 61.52
√ √ 61.59
√ √ 62.58
√ √ 61.70
√ √ 61.96

Table 3. Ablation experiment for value selection about η in Eq. (4)
View table
Table 3. Ablation experiment for value selection about η in Eq. (4)
Models η=5 η=10 η=15 η=20
mAP /% 62.13 62.37 62.58 62.55

Table 4. Comparison of ablation experimental results

View table

Table 4. Comparison of ablation experimental results

Baseline	FR-FPN	ATT-ORPN	PA-head	SWA	mAP /%	mAP variation /%	Recall /%	f /（frame·s^-1）
√					59.54		—	—
√	√				63.03	↑3.49	70.68	51.3
√		√			62.36	↑2.82	69.90	29.8
√			√		62.58	↑3.04	70.31	42.6
√	√	√	√		63.83	↑4.29	71.16	18.9
√	√	√	√	√	64.49	↑4.95	71.35	12.8

Table 5. Comparison of precision of different network models on Dior-R dataset

View table

Table 5. Comparison of precision of different network models on Dior-R dataset

Model	RetinaNet-O^［22］	Rotated RPN^［15］	Gliding Vertex^［23］	ROI Transformer^［11］	AOPG^［19］	PFR-Rotate	PFR-Rotate（ms）
mAP	57.55	59.54	60.06	63.87	64.41	64.49	65.97
APL	61.49	62.79	65.35	63.34	62.39	64.25	64.54
APO	28.52	26.8	28.87	37.88	37.79	39.78	41.2
BF	73.57	71.72	74.96	71.78	71.62	73.39	75.5
BC	81.17	80.91	81.33	87.53	87.63	82.13	83.55
BR	23.98	34.2	33.88	40.68	40.9	40.98	41.69
CH	72.54	72.57	74.31	72.6	72.47	73.89	75.54
DAM	19.94	18.95	19.58	26.86	31.08	29.98	31.35
ETS	72.39	66.45	70.72	78.71	65.42	70.91	73.02
ESA	58.2	65.75	64.7	68.09	77.99	77.56	78.82
GF	69.25	66.63	72.3	68.96	73.2	79.18	79.4
GTF	79.54	79.24	78.68	82.74	81.94	81.66	83.55
HA	32.14	34.95	34.22	47.71	42.32	37.39	40.19
OP	44.87	48.79	74.64	55.61	54.45	50.62	53.56
SH	77.71	81.14	80.22	81.21	81.17	82.18	83.38
STA	67.57	64.34	69.26	78.23	72.69	76.24	78.21
STO	61.09	71.21	61.13	70.26	71.31	72.68	74.56
TC	81.46	81.44	81.49	81.61	81.49	81.53	82.05
TS	47.33	47.31	44.76	54.86	60.04	57.52	58.75
VE	38.01	50.46	47.71	43.27	52.38	50.57	51.64
WM	60.24	65.21	65.04	65.52	69.99	67.51	69.64

Table 6. Comparison of directional and directed labeling results of Dior-R dataset

View table

Table 6. Comparison of directional and directed labeling results of Dior-R dataset

Directional labeling			Directed labeling
Model	Backbone	mAP /%	Model	Backbone	mAP /%
SSD^［24］	VGG16	58.6	Retinanet-O^［22］	Resnet50	57.55
YOLO V3^［25］	Darknet53	57.1	Rotated RPN^［15］	Resnet50	59.54
Cascade R-CNN^［26］	Resnet50	60.5	Gliding Vertex^［23］	Resnet50	60.06
SA-Cascade^［27］	Resnet50	62.1	PRF-Rotate	Resnet50	64.49

Table 7. Comparison of test results of different network models on HRSC2016 dataset

View table

Table 7. Comparison of test results of different network models on HRSC2016 dataset

Model	Backbone	mAP（07）/%	mAP（12）/%
RetinaNet-O^［22］	Resnet101	89.18	95.21
Rotated RPN^［15］	Resnet50	79.08	85.64
Gliding Vertex^［23］	Resnet101	88.20	—
ROI Transformer^［11］	Resnet101	86.20	—
S²A-Net^［25］	Resnet101	90.17	95.01
AOPG^［19］	Resnet101	90.34	96.22
Oriented R-CNN^［28］	Resnet50	90.40	96.50
Oriented RepPoints^［29］	Resnet50	90.38	97.26
AFO-RPN^［30］	Resnet101	90.45	—
PFR-Rotate	Resnet50	90.83	97.35

Tools

Get Citation

Copy Citation Text

Youwei Wang, Ying Guo, Xiangying Shao, Jiyu Wang, Zhengwei Bao. Oriented Object Detection in Remote Sensing Images Based on Feature Recombination[J]. Acta Optica Sinica, 2024, 44(6): 0628001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites