Remote Sensing Image Detection Algorithm Based on Multi-receptive Field and Dynamic Feature Refinement

Jun Huang; Ying Guo; Shu Yan

doi:10.3788/LOP240932

Laser & Optoelectronics Progress, Volume. 61, Issue 22, 2228004(2024)

Remote Sensing Image Detection Algorithm Based on Multi-receptive Field and Dynamic Feature Refinement

Jun Huang^1,2、*, Ying Guo^1,2, and Shu Yan^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(15)

Fig. 1. Network structure of YOLOv7^[16]

Download full size

Fig. 2. Network structure of DF-YOLOv7

Download full size

Fig. 3. Schematic diagram of detection layer transformation^[16]

Download full size

Fig. 4. Three structures of DRES module. (a) DRES_1; (b) DRES_2; (c) DRES_3

Download full size

Fig. 5. MRELAN module

Download full size

Fig. 6. EMA attention mechanism

Download full size

Fig. 7. Feature refinement module DCR

Download full size

Fig. 8. Comparison of mAP, parameters, and FPS between DF-YOLOv7 and various networks

Download full size

Fig. 9. Comparison of heat maps between YOLOv7 and DF-YOLOv7. (a) Original image; (b) YOLOv7 heat image; (c) DF-YOLOv7 heat image; (d) YOLOv7 detection image; (e) DF-YOLOv7detection image

Download full size

Fig. 10. Comparison of detection effect between YOLOv7 and DF-YOLOv7 on VisDrone images

Download full size

Table 1. Ablation experiments of DRES module on the VisDrone verification set
View table
Table 1. Ablation experiments of DRES module on the VisDrone verification set
Model mAP0.5 /% Parameters /10⁶ GFLOPs FPS /（frame/s）
LYS+DRES_1 51.2 17.8 116.0 92.4
LYS+DRES_2 51.5 17.5 114.4 97.3
LYS+SRES_3 51.3 17.5 114.5 95.7

Table 2. Ablation of attention mechanisms on the VisDrone validation set
View table
Table 2. Ablation of attention mechanisms on the VisDrone validation set
Model mAP0.5 /% Parameters /10⁶ GFLOPs
Base+CBAM
51.5
51.5
17.5
18.0
114.4
118.2
+SE 51.2 18.0 118.2
+SimAM 51.5 18.0 118.2
+GAM 51.4 18.1 119.0
+EMA 51.7 18.0 118.3

Table 3. Ablation experiments on the VisDrone validation set

View table

Table 3. Ablation experiments on the VisDrone validation set

Model	mAP0.5 /%	Parameters /10⁶	GFLOPs	FPS /（frame/s）
YOLOv7	49.0	37.3	103.3	98.4
+P₂	50.2	37.1	117.1	80.5
+P₂-P₅	50.7	17.7	115.3	102.1
+P₂-P₅+MRELAN	51.7	18.0	118.3	93.6
+P₂-P₅+MRELAN+DCR	52.1	18.3	130.6	69.5
+P₂-P₅+MRELAN+DCR+Wise IoU v3	52.3	18.3	130.6	71.3

Table 4. Comparison of the effect of DF-YOLOv7 and other detectors on VisDrone verification set

View table

Table 4. Comparison of the effect of DF-YOLOv7 and other detectors on VisDrone verification set

Model	Precision /%										mAP0.5/%	Parameters /10⁶	FPS/（frame/s）
Model	Pede	Peop	Bicy	Car	Van	Truck	Tricy	Awn	Bus	Motor	mAP0.5/%	Parameters /10⁶	FPS/（frame/s）
Faster R-CNN	21.4	15.6	6.7	51.7	29.5	19.0	13.1	7.7	31.4	20.7	21.7	41.2	‒
Cascade-RCNN^［31］	37.9	27.7	13.3	74.3	44.6	34.7	24.6	13.2	52.4	38.3	36.1	42.7	‒
HRDNet^［32］	56.7	45.1	27.7	82.6	51.3	43.0	37.6	18.8	58.9	56.4	47.8	62.4	3.7
YOLOv4^［33］	24.8	12.6	8.6	64.3	22.4	22.7	11.4	7.6	44.3	21.7	24.0	64.0	39.3
YOLOv5s	40.3	32.1	9.9	72.7	33.4	26.4	18.5	11.6	39.0	38.1	32.2	7.2	90.1
MC-YOLOv5^［34］	40.9	33.5	10.7	74.2	37.2	27.9	18.7	12.4	43.4	38.9	33.8	38.2	57.1
YOLOv6m^［35］	34.5	25.7	5.6	75.3	42.5	31.3	22.0	14.0	48.3	37.0	33.6	34.8	56.0
YOLOv7	57.4	48.2	24.5	84.8	50.2	45.6	38.5	18.6	63.2	59.2	49.0	37.2	98.0
YOLOv8	42.7	32.0	12.4	79.1	44.0	36.5	28.1	15.9	57.0	44.9	39.3	11.1	120.5
UAV-YOLOv8s^［36］	56.8	44.9	18.8	85.8	50.8	39.0	33.3	19.7	64.3	56.2	47.0	10.3	52.6
DF-YOLOv7	62.1	51.4	28.9	86.7	54.1	48.0	41.1	21.6	66.1	62.1	52.3	18.3	71.3

Table 5. Comparison of DF-YOLO and other algorithms on DOTA dataset

View table

Table 5. Comparison of DF-YOLO and other algorithms on DOTA dataset

Model	Precision
Model	R³Det^［37］	RSDet^［38］	SCRDet^［39］	ICN^［40］	RADet^［41］	YOLOv5	YOLOv7	Proposed
mAP	73.8	74.1	72.6	68.2	69.1	71.3	76.9	79.2
PL	89.5	90.1	90.0	81.4	79.5	89.3	89.8	89.7
BD	81.2	82.0	80.7	74.3	77.0	76.4	83.6	80.9
BR	50.5	53.8	52.1	47.7	48.1	47.3	50.2	58.1
GTF	66.1	68.5	68.4	70.3	65.8	61.2	66.7	78.5
SV	71.0	70.2	68.4	64.9	65.5	71.3	78.3	79.1
LV	78.7	78.7	60.3	67.8	74.4	74.0	83.1	83.9
SH	78.2	73.6	72.4	70.0	68.9	78.6	87.3	86.3
TC	90.8	91.3	90.9	90.8	89.7	89.8	90.7	90.7
BC	85.3	87.1	87.9	79.1	78.1	82.2	85.4	83.6
ST	84.2	84.7	86.9	78.3	75.0	81.4	85.7	85.7
SBF	61.8	64.3	65.0	53.6	49.9	60.9	65.5	67.2
RA	63.8	68.2	66.7	62.9	64.6	63.9	64.9	67.0
HA	68.2	66.2	66.3	67.0	66.1	65.2	67.5	78.1
SP	70.0	69.3	68.2	64.2	71.6	68.4	78.2	81.3
HC	67.2	63.7	65.2	50.2	62.2	60.1	76.0	77.3

Tools

Get Citation

Copy Citation Text

Jun Huang, Ying Guo, Shu Yan. Remote Sensing Image Detection Algorithm Based on Multi-receptive Field and Dynamic Feature Refinement[J]. Laser & Optoelectronics Progress, 2024, 61(22): 2228004

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites