Target Detection in Remote Sensing Images Based on Improved Cascade Algorithm

Youwei Wang; Ying Guo; Xiangying Shao

doi:10.3788/AOS202242.2428004

Acta Optica Sinica, Volume. 42, Issue 24, 2428004(2022)

Target Detection in Remote Sensing Images Based on Improved Cascade Algorithm

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

¹Jiangsu Collaborative Innovation Center of Atmospheric Environment and Equipment Technology (CICAEET), 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(14)

Fig. 1. Framework of Cascade R-CNN

Download full size

Fig. 2. Framework of SA-Cascade

Download full size

Fig. 3. Framework of R-FPN

Download full size

Fig. 4. ASPP connection mode

Download full size

Fig. 5. Comparison of anchor generation methods. (a) Common RPN anchor generation method;（b）LA-RPN anchor generation method

Download full size

Fig. 6. Framework of LA-RPN

Download full size

Fig. 7. Structure comparison diagram of detection head. (a) Full connection layer structure detection head;（b）db-head detection head

Download full size

Fig. 8. Framework of db-head

Download full size

Fig. 9. Comparison of SA-Cascade and other algorithms on detection effect. (a) Faster R-CNN+ResNet50;(b) Cascade R-CNN+ResNet50; (c) SA-Cascade+ResNet50

Download full size

Table 1. Comparison of ablation experimental results

View table

Table 1. Comparison of ablation experimental results

Cascade framework	R-FPN	LA-RPN	db-head	mAP /%	Increased percentage	Recall /%	Detection rate /（frame·s^-1）
				89.84		93.10	47.3
√				90.17	0.33	92.60	18.4
√	√			91.12	1.28	92.64	16.5
√		√		91.78	1.94	94.55	13.0
√			√	90.67	0.83	94.23	9.0
√	√	√		92.73	2.89	94.92	15.4
√	√		√	91.61	1.77	94.45	6.7
√		√	√	91.64	1.80	94.49	16.6
	√	√	√	92.10	2.26	94.76	13.0
√	√	√	√	93.11	3.27	94.82	10.4

Table 2. Comparison of results of different backbones

View table

Table 2. Comparison of results of different backbones

Model	Network	mAP /%	AP for IoU of 0.5 /%	Recall /%
Cascade R-CNN^［10］	ResNet50	84.90
Cascade R-CNN （ours）	ResNet50	90.17	90.20	92.17
	ResNet101	90.66	90.70	92.60
	ResNet152	91.31	91.30	92.97
SA-Cascade	ResNet50	93.11	93.10	94.82
	ResNet101	94.62	94.60	96.75
	ResNet152	94.98	94.96	96.92

Table 3. Accuracy comparison of different size targets unit: %
View table
Table 3. Accuracy comparison of different size targets unit: %
Model mAP mAP_s mAP_m mAP_l
Cascade R-CNN 90.17 44.40 83.50 84.40
SA-Cascade 93.11 58.70 86.30 87.30

Table 4. Comparison of accuracy of this algorithm and other algorithms on TGRS-HRRSD- Dataset unit: %

View table

Table 4. Comparison of accuracy of this algorithm and other algorithms on TGRS-HRRSD- Dataset unit: %

Model	YOLO V3^*［26］	Faster R-CNN^▲［11］	Faster R-CNN^［11］	MS HEMN^［27］	Cascade R-CNN^［10］	Cascade R-CNN（ours）	SA-Cascade	SA-Cascade（ms）
mAP	71.59	81.50	83.40	85.90	84.90	90.17	93.11	94.46
APL	90.7	90.8	90.9	90.9	90.9	98.1	99.2	99.2
BD	76.4	86.9	87.4	88.5	88.7	93.3	95.4	95.7
BC	66.3	47.9	65.0	74.0	68.7	68.9	72.4	80.4
BR	75.2	85.5	84.3	88.3	88.5	94.1	97.5	97.5
CR	75.0	88.6	88.7	88.7	88.8	92.9	96.8	97.2
GTF	89.4	90.6	90.7	90.8	90.8	98.0	99.0	99.1
HA	45.9	89.4	89.1	88.8	90.0	97.8	97.5	98.1
PL	42.3	69.3	64.1	68.1	63.0	69.4	76.7	81.8
SH	73.6	88.5	87.5	89.6	89.8	94.2	95.5	95.7
ST	77.0	88.7	89.8	89.9	90.0	95.2	96.5	97.9
TJ	59.4	75.1	70.8	78.7	74.3	80.3	89.7	91.9
TC	86.2	80.7	89.9	90.4	90.3	92.2	93.5	95.4
VE	73.3	84.0	89.4	90.2	90.3	97.8	98.2	98.1

Table 5. Comparison between results of this algorithm and other algorithms on VisDrone-DET dataset

View table

Table 5. Comparison between results of this algorithm and other algorithms on VisDrone-DET dataset

Model	mAP	AP50	AP75	AR100
CornerNet^［28］	17.41	34.12	15.78	24.37
Light-RCNN^［29］	16.53	32.78	15.13	23.09
FPN^［30］ DetNet59^［31］	16.51 15.26	32.20 29.23	14.91 14.34	20.72 20.87
RefineDet^［32］	14.90	28.76	14.08	18.13
RetinaNet^［17］	11.81	21.37	11.62	5.31
Cascade R-CNN^［10］	16.09	31.91	15.01	21.37
SA-Cascade	25.80	41.70	27.20	34.90

Tools

Get Citation

Copy Citation Text

Youwei Wang, Ying Guo, Xiangying Shao. Target Detection in Remote Sensing Images Based on Improved Cascade Algorithm[J]. Acta Optica Sinica, 2022, 42(24): 2428004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites