Lightweight Target Detection Network Integrating Scene Context

Tingting Liu; Hua Miao; Lin Li; Yang Xiang; Qi Li; Qi Meng

doi:10.3788/LOP202158.2010005

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 20 >Page 2010005 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 20, 2010005 (2021)

Lightweight Target Detection Network Integrating Scene Context

Tingting Liu, Hua Miao^*, Lin Li, Yang Xiang, Qi Li, and Qi Meng

Author Affiliations

School of Opto-Electronic Engineering, Changchun University of Science and Technology, Changchun, Jilin 130022, China

show less

DOI: 10.3788/LOP202158.2010005 Cite this Article Set citation alerts

Tingting Liu, Hua Miao, Lin Li, Yang Xiang, Qi Li, Qi Meng. Lightweight Target Detection Network Integrating Scene Context[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010005 Copy Citation Text

show less

Fig. 1. Structure diagram of SS-YOLOv3

Download full size

Fig. 2. Structure diagram of MobileNet-YOLOv3

Download full size

Fig. 3. Test results of YOLOv3

Download full size

Fig. 4. Scene context module

Download full size

Fig. 5. Typical aerial images. (a) Large scale variation; (b) complex background; (c) most small goals

Download full size

Fig. 6. Comparison of visual test results and enlarged view of some areas. (a) YOLOv3 test results; (b) SS-YOLOv3 test results

Download full size

Fig. 7. Visualization results of SS-YOLOv3

Download full size

Weight decay	Batch size	Learning rate	Momentum	Maximum iteration
0.0005	16	0.001 (M<1000)0.0001(M>1000)	0.9	10000

Table 1. Hyperparameters of pre-trained model

Algorithm	Backbone	mAP /%	FPS /(frame·s^-1)
Fast R-CNN	VGG16	70.0	6
Faster R-CNN	VGG16	75.8	8
YOLO		64.3	18
SSD₃₀₀	VGG16	73.4	39
SSD₅₀₀	VGG16	74.9	22
YOLOv2	Darknet-19	74.3	34
YOLOv3	Darknet-53	75.9	27.6
SS-YOLOv3	MobileNetV3	84.3	33.4

Table 2. Results of different target detection algorithms

Class	YOLOV3	SS-YOLOv3	Subtract
Aeroplane	74.6	84.9	10.3
Car	75.9	84.0	8.1
Boat	75.7	85.6	9.9
Cycle	76.3	82.4	6.1
House	74.3	84.6	10.3
Electrocar	76.9	82.7	5.8
Person	77.6	85.9	8.3

Table 3. Detection results of YOLOv3 and SS-YOLOv3 unit: %

YOLOv3	MobileNetV3	Scene SRU	CIOU	mAP /%	FPS /(frame·s^-1)
√				75.9	27.6
√	√			74.2	36.4
√	√	√		83.5	33.4
√	√	√	√	84.3	33.4

Table 4. Results of ablation learning

Algorithm	Pedestrian	Person	Bicycle	Car	Van	Truck	Tricyle	Awn	Bus	Motor
YOLOv3	20.75	8.04	7.43	45.82	30.56	22.89	13.04	9.27	33.84	11.46
SS-YOLOv3	22.45	7.93	9.65	51.03	34.54	24.13	15.02	9.13	36.89	13.73

Table 5. Detect results of YOLOv3 and SS-YOLOv3 unit: %

Tingting Liu, Hua Miao, Lin Li, Yang Xiang, Qi Li, Qi Meng. Lightweight Target Detection Network Integrating Scene Context[J]. Laser & Optoelectronics Progress, 2021, 58(20): 2010005

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information