Lightweight and high-precision object detection algorithm based on YOLO framework

Table 1. Comparison experiment of lightweight backbone network
View table
View in Article
Table 1. Comparison experiment of lightweight backbone network
算法参数量/M GFLOPs mAP@0.5/%
ShuffleNetv2-0.5 2.57 4.9 32.4
Ghost-0.5 5.32 13.2 36.6
MobileNetv2-0.5 1.43 4.0 38.9
MobileNetv3-0.5 1.29 3.7 40.6
E-MobileNetv3-0.5 1.27 3.8 41.6

Table 2. Verification experiment of SSH embedding mode
View table
View in Article
Table 2. Verification experiment of SSH embedding mode
算法参数量/M GFLOPs mAP@0.5/%
FPN 0.6 2.4 33.65
FPN+SSH 1.27 3.6 34.72
SSH+FPN 1.29 4.0 40.36

Table 3. Experimental results of different activation functions in Neck part
View table
View in Article
Table 3. Experimental results of different activation functions in Neck part
Relu Hard-Swish Mish mAP@0.5/%
√ 40.36
√ 40.44
√ 40.60

Table 4. Ablation experiments

View table

View in Article

Table 4. Ablation experiments

Focus	ECA	CBAM	GIOU	DIOU	CIOU	SIOU	SSH	PaFPN	Soft-NMS	YOLOX	参数量/M	mAP@0.5/%
					√		√				1.29	40.60
√					√		√				1.291	40.66
√		√			√		√				1.28	41.7
√	√				√		√				1.27	41.6
√	√				√						0.6	33.65
√	√		√				√				1.27	37.28
√	√			√			√				1.27	38.8
√	√				√		√	√			2.35	45.2
√	√				√		√		√		1.27	45.6
√	√				√		√			√	2.09	47.8
√	√				√		√		√	√	2.09	52.8
√	√					√	√			√	2.09	50.6
√	√					√	√		√	√	2.09	54.5

Table 5. Comparative experiments on MS COCO

View table

View in Article

Table 5. Comparative experiments on MS COCO

算法	Backbone	参数量/M	GFLOPs	mAP@0.5/%
YOLOv4-tiny	CSPDarkNet53	5.78	16.4	44.6
YOLOv5s	CSPDarkNet53	6.73	16.4	54.8
YOLOXs	CSPDarkNet53	8.51	26.6	58.2
YOLOXs	MobileNetv3	4.7	16.1	53.2
YOLOv5s	MobileNetv3	3.24	6.5	50.3
Faster R-CNN	Resnet50	41.75	-	57.2
Mask R-CNN	Resnet50+FPN	44.40	-	58.5
Cascade R-CNN	Resnet50+FPN	69.39	-	58.3
SSD	VGG16	36.04	-	49.1
SSD	MobileNetv2	4.23	-	35.8
FCOS	Resnet50	32.30	-	55.3
CenterNet	Resnet18	14.4	-	46.2
EfficientNet-b3	-	19.98	-	59.8
CentripetalNet	HourglassNet	205.76	-	61.9
Ours	E-MobileNetv3	3.0	11.4	58.6

Table 6. Comparative experiments on the UA-DETRAC
View table
View in Article
Table 6. Comparative experiments on the UA-DETRAC
算法 AP@0.5/% FPS
Car Bus Van Others
YOLOv3_SPP 87.4 86.9 59 57.5 204
YOLOv5s 84.6 84.7 56.5 39.4 303
YOLOXs 85.8 86 57 43.6 250
Ours 86.6 86.3 61.2 47.9 286

Tools

Get Citation

Copy Citation Text

Xin-chuan FAN, Chun-mei CHEN. Lightweight and high-precision object detection algorithm based on YOLO framework[J]. Chinese Journal of Liquid Crystals and Displays, 2023, 38(7): 945

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites