Improved Faster R-CNN Target Detection Algorithm Based on Attention Mechanism and Soft-NMS

Fengsui Wang; Qisheng Wang; Jingang Chen; Furong Liu

doi:10.3788/LOP202158.2420001

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 24 >Page 2420001 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 24, 2420001 (2021)

Improved Faster R-CNN Target Detection Algorithm Based on Attention Mechanism and Soft-NMS

Fengsui Wang^{1、2、3、*}, Qisheng Wang^1、2、3, Jingang Chen^1、2、3, and Furong Liu^1、2、3

Author Affiliations

¹School of Electrical Engineering, Anhui Polytechnic University, Wuhu, Anhui 241000, China

²Anhui Key Laboratory of Detection Technology and Energy Saving Devices, Wuhu, Anhui 241000, China

³Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment, Ministry of Education, Wuhu, Anhui 241000, China

show less

DOI: 10.3788/LOP202158.2420001 Cite this Article Set citation alerts

Fengsui Wang, Qisheng Wang, Jingang Chen, Furong Liu. Improved Faster R-CNN Target Detection Algorithm Based on Attention Mechanism and Soft-NMS[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2420001 Copy Citation Text

show less

Fig. 1. Faster R-CNN target detection model

Download full size

Fig. 2. Structure of improved convolutional attention mechanism

Download full size

Fig. 3. Improved Faster R-CNN model

Download full size

Fig. 4. Comparison of false detection rate. (a) Error detection rate statistics of original algorithm; (b) error detection rate statistics of improved algorithm

Download full size

Fig. 5. Image detection comparison. (a) Detection results of original algorithm; (b) detection results of improved algorithm

Download full size

Dataset	Training		Test
Dataset	Image	Object	Image	Object
VOC2007	5011	12608	4952	12032
VOC2012	11540	27450	0	0
Total	16551	40058	4952	12032

Table 1. Statistics of training and test data in VOC2007 and VOC2012 datasets

Serial number	Model	Backbone	Channel attention	Spatial attention	Dimension reduction	NMS	Soft-NMS
0	FR^[11]	VGG-16	--	--	--	√	--
1	FR	ResNet-50	--	--	--	√	--
2	+Soft	ResNet-50	--	--	--	--	√
3	+SE	ResNet-50	√	--	√	√	--
4	+ECA	ResNet-50	√	--	--	√	--
5	+CBAM	ResNet-50	√	√	√	√	--
6	Ours1	ResNet-50	√	√	--	√	--
7	Ours2	ResNet-50	√	√	--	--	√

Table 2. Comparison of experimental model structure

Category	FR		Ours1		Variation		Category		SSD		SSD+		Variation		Category		YOLOv4		YOLOv4+		Variation
Cat	87.2		89.2		+2.0		Cat		89.7		89.6		-0.1		Cat		90.3		90.2		-0.1
Car	86.6		85.4		-1.2		Car		84.9		85.1		+0.2		Car		94.8		95.1		+0.3
Horse	84.9		86.6		+1.7		Horse		89.1		89.4		+0.3		Horse		91.7		91.0		-0.7
Dog	83.6		85.6		+2.0		Dog		85.7		86.1		+0.4		Dog		87.1		88.9		+1.8
Bus	80.3		85.3		+5.0		Bus		84.2		83.9		-0.3		Bus		91.3		92.6		+1.3
Train	82.5		82.5		0		Train		87.1		87.2		+0.1		Train		93.1		92.6		-0.5
Motorbike	83.5		81.0		-2.5		Motorbike		84.2		84.5		+0.3		Motorbike		92.1		92.1		0
Bicycle	79.6		80.9		+1.3		Bicycle		86.7		87.1		+0.4		Bicycle		90.7		91.1		+0.5
Person	79.8		79.7		-0.1		Person		81.4		81.2		-0.2		Person		91.2		91.1		-0.1
Aeroplane	78.1		77.6		-0.5		Aeroplane		76.2		77.5		+1.3		Aeroplane		87.9		90.4		+2.5
Category		FR		Ours1		Variation		Category		SSD		SSD+		Variation		Category		YOLOv4		YOLOv4+		Variation
Sheep		75.2		76.9		+1.7		Sheep		75.3		77.1		+1.8		Sheep		87.4		87.7		+0.3
Bird		74.1		75.3		+1.2		Bird		75.4		75.0		-0.4		Bird		87.4		86.0		-1.4
Cow		74.7		75.1		+0.4		Cow		78.5		79.6		+1.1		Cow		91.5		92.1		+0.6
Tvmonitor		73.3		72.5		-0.8		Tvmonitor		76.6		76.7		+0.1		Tvmonitor		89.0		90.1		+1.1
Diningtable		72.3		73.0		+0.7		Diningtable		76.3		80.1		+3.8		Diningtable		80.9		81.9		+1.0
Sofa		70.4		75.2		+4.8		Sofa		78.0		80.4		+1.6		Sofa		77.1		79.2		+2.1
Boat		65.7		66.4		+1.3		Boat		67.3		66.4		-0.9		Boat		75.4		79.0		+3.6
Chair		54.2		53.4		-0.8		Chair		59.7		61.2		+0.5		Chair		73.2		71.9		-1.3
Bottle		52.1		53.6		+1.5		Bottle		49.2		48.8		-0.4		Bottle		82.1		80.8		-1.3
Pottedplant		46.0		44.4		-1.6		Pottedplant		47.2		47.6		+0.4		Pottedplant		60.0		60.6		+0.6
mAP		74.2		75.0		+0.8		mAP		76.6		77.2		+0.6		mAP		85.7		86.2		+0.5

Table 3. Results of validation experiment of improved attention mechanism unit: %

Model	Backbone	Input image size /(pixel×pixel)	mAP /%
FR^[11]	VGG-16	~600×1000	73.2
FR	ResNet-50	~600×1000	74.2
+Soft	ResNet-50	~600×1000	74.5
+SE	ResNet-50	~600×1000	75.2
+ECA	ResNet-50	~600×1000	74.8
+CBAM	ResNet-50	~600×1000	74.8
Ours1	ResNet-50	~600×1000	75.0
Ours2	ResNet-50	~600×1000	75.9

Table 4. MAP comparison of experimental models on VOC2007 test set

Category	FR^[11]	FR	+Soft	+SE	+ECA	+CBAM	Ours1	Ours2
Cat	86.4	87.2	87.4	89.9	88.6	87.4	89.2	90.1
Car	84.7	86.6	87.1	85.6	85.2	85.1	85.4	86.1
Horse	84.6	84.9	85.2	86.6	86.1	86.6	86.6	86.8
Dog	84.8	83.6	84.0	87.8	84.1	82.6	85.6	86.8
Bus	83.1	80.3	80.3	83.1	84.9	84.7	85.3	85.3
Train	83.0	82.5	84.2	84.1	81.9	83.0	82.5	84.4
Motorbike	77.5	83.5	82.9	82.5	83.4	80.8	81.0	82.5
Bicycle	79.0	79.6	80.4	81.5	83.3	81.5	80.9	82.6
Person	76.7	79.8	80.5	81.0	79.6	80.8	79.7	81.3
Aeroplane	76.5	78.1	78.9	75.9	78.4	76.9	77.6	78.7
Sheep	73.6	75.2	74.4	73.0	73.5	70.5	76.9	77.3
Bird	70.9	74.1	74.2	76.0	75.9	75.8	75.3	76.8
Cow	81.9	74.7	75.0	77.5	74.0	77.2	75.1	75.9
Tvmonitor	72.6	73.3	74.4	72.1	73.0	75.4	72.5	73.1
Diningtable	65.7	72.3	72.0	71.7	72.5	72.4	73.0	73.6
Sofa	73.9	70.4	70.9	73.5	71.1	71.6	75.2	76.2
Boat	65.5	65.7	65.9	66.1	66.4	66.1	66.4	66.7
Chair	52.0	54.2	54.1	53.6	53.7	54.1	53.4	54.1
Bottle	52.1	52.1	52.6	59.1	54.8	56.0	53.6	54.0
Pottedplant	38.8	46.0	46.0	44.2	45.9	45.6	44.4	46.1
mAP	73.2	74.2	74.5	75.2	74.8	74.8	75.0	75.9

Table 5. AP comparison of 20 target classes unit: %

Algotithm	Backbone	mAP /%
Ref. [11]	VGG-16	73.2
Ref. [22]	ResNet-101	76.4
Ref. [23]	ResNet-50	74.4
Ref. [24]	ResNet-101	74.4
YOLO^[6]	Darknet	63.4
SSD300^[7]	VGG-16	74.3
Ref. [25]	DiCENet	68.4
Ref. [26]	ResNet-50	74.4
Ours2	ResNet-50	75.9

Table 6. mAP comparison of different algorithms

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information