Attention Mechanism-Based Improved Lightweight Target Detection Algorithm

Table 1. Experimental dataset details
View table
Table 1. Experimental dataset details
Dataset Class of number Train set Test set
PASCAL VOC 20 15412 1713
MS COCO 80 105539 11727

Table 2. Comparison of three improved methods
View table
Table 2. Comparison of three improved methods
Method PASCAL VOC
CSPdarknet53-s ATT-block ATT-FPN mAP Recall
× × × 67 75.6
√ × × 73.3 79.1
√ √ × 84.5 89.6
√ √ √ 87.9 91.2

Table 3. VOC2012 test comparison
View table
Table 3. VOC2012 test comparison
Method Backbone mAP /% Recall /% FPS
Faster R-CNN Resnet50 75.7 78.3 9
YOLOv4 CSPdarknet53 86 89.5 32
YOLOv4-tiny CSPdarknet53s 67 75.6 106
YOLOv4s CSPdarknet53-s 87.9 91.2 90

Table 4. Comparison of AP of each category in VOC2012

View table

Table 4. Comparison of AP of each category in VOC2012

Category	Faster R-CNN	YOLOv4	YOLOv4-tiny	YOLOv4s	Category	Faster R-CNN	YOLOv4	YOLOv4-tiny	YOLO v4s
Aeroplane	82.8	92.8	67.9	92.8	Table	80.4	79.5	74.5	88.6
Bicycle	71.0	80.6	67.5	85.4	Dog	90.4	97.6	85.6	94.3
Bird	90.6	94.7	78.3	93.9	Horse	60.8	51.9	64.9	87.8
Boat	71.5	82.7	54.7	78.5	Mbike	70.0	64.5	64.8	89.1
Bottle	50.0	76.1	31.6	77.4	Person	77.2	81.6	68.2	86.5
Bus	89.8	96.3	90.5	96.8	Plant	56.9	82.8	50.0	72.3
Car	63.3	81.1	53.1	82.5	Sheep	94.7	95.9	58.6	94.8
Cat	94.6	96.7	88.2	97.8	Sofa	79.4	89.1	74.7	92.0
Chair	51.1	67.6	46.4	77.8	Train	83.0	96.8	87.0	97.2
Cow	83.5	95.0	67.2	91.6	Tv	73.4	76.5	66.6	87.1

Table 5. MS COCO test comparison

View table

Table 5. MS COCO test comparison

Method	Backbone	AP	AP₅₀	AP₇₅	AP_S	AP_M	AP_L	FPS
YOLOv4-tiny	CSPDarknet53s	18.1	37.0	15.4	5.6	20.6	29.3	106
Efficientdet-b0	Efficientnet	23.3	39.1	26.1	8.1	26.5	38.5	22
YOLOv4s-1	CSPDarknet53-s	19.3	38.9	20.2	6.2	22.3	33.5	108
YOLOv4s-2	CSPDarknet53s	22.5	43.9	26.5	8.3	25.6	40.8	89
YOLOv4s	CSPDarknet53-s	25.0	46.1	27.2	8.6	29.2	45.7	90

Tools

Get Citation

Copy Citation Text

Mei Jin, Yihui Li, Liguo Zhang, Zijian Ma. Attention Mechanism-Based Improved Lightweight Target Detection Algorithm[J]. Laser & Optoelectronics Progress, 2023, 60(4): 0415008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites