Faster R-CNN Target-Detection Algorithm Fused with Adaptive Attention Mechanism

Qisheng Wang; Fengsui Wang; Jingang Chen; Furong Liu

doi:10.3788/LOP202259.1215016

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 12 >Page 1215016 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 12, 1215016 (2022)

Faster R-CNN Target-Detection Algorithm Fused with Adaptive Attention Mechanism

Qisheng Wang^1、2、3, Fengsui 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 241000, Anhui , China

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

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

show less

DOI: 10.3788/LOP202259.1215016 Cite this Article Set citation alerts

Qisheng Wang, Fengsui Wang, Jingang Chen, Furong Liu. Faster R-CNN Target-Detection Algorithm Fused with Adaptive Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215016 Copy Citation Text

show less

Fig. 1. Structure diagram of MA

Download full size

Fig. 2. Comparison diagram of test results. (a) Original algorithm detection results; (b) Detection results of improved algorithm based on MA

Download full size

Layer name	Output size	ResNet-50	ResNet-50_MA*	ResNet-50_MA
Conv1	112×112	7×7，64，stride 2
Attention 0	112×112			MA
Conv2_x	56×56	3×3 max pool，stride 2
Conv2_x	56×56	$\{\begin{matrix} 1 \times 1, 64 \\ 3 \times 3, 64 \\ 1 \times 1, 256 \end{matrix}\} \times 3$
Conv3_x	28×28	$\{\begin{matrix} 1 \times 1, 128 \\ 3 \times 3, 128 \\ 1 \times 1, 512 \end{matrix}\} \times 4$
Conv4_x	14×14	$\{\begin{matrix} 1 \times 1, 256 \\ 3 \times 3, 256 \\ 1 \times 1, 1024 \end{matrix}\} \times 6$
Conv5_x	7×7	$\{\begin{matrix} 1 \times 1, 512 \\ 3 \times 3, 512 \\ 1 \times 1, 2048 \end{matrix}\} \times 3$
Attention 1	7×7		MA*
	1×1	Average pool，fc，softmax

Table 1. Comparison of modified ResNet-50 structure based on MA

Name	Model number
System	Windows 10
CPU	Core i9-10900 @ 3.7 GHz
GPU	Nvidia GeForce RTX 2080Ti（11 GB）
Framework	Pytorch 1.2.0
Language	Python

Table 2. Experimental environment configuration

Data	Trainval		Test
Data	Images	Objects	Images	Objects
VOC2007	5011	12608	4952	12032
VOC2012	11540	27450	0	0
Total	16551	40058	4952	12032

Table 3. PASCAL VOC dataset training and test data statistics

Category	FR	FR+MA*	FR+MA
mAP	74.25	75.01^+0.76	75.16^+0.91
Cat	87.23	87.31^+0.08	89.28^+2.05
Car	86.60	86.33^-0.27	85.85^-0.75
Horse	84.87	86.14^+1.27	86.50^+1.63
Dog	83.62	82.87^-0.75	84.29^+0.67
Bus	80.29	82.55^+2.26	83.45^+3.16
Train	82.49	83.12^+0.63	82.96^+0.47
Motorbike	83.48	83.62^+0.14	83.07^-0.41
Bicycle	79.56	83.14^+3.58	82.72^+3.16
Person	79.75	80.09^+0.34	80.38^+0.63
Aeroplane	79.75	76.09^-3.66	74.28^-5.47
Sheep	75.22	78.26^+3.04	73.20^-2.02
Bird	74.14	76.45^+2.31	77.53^+3.39
Cow	74.73	78.78^+4.05	74.70^-0.03
Tvmonitor	74.34	73.33^-1.01	73.09^-1.25
Diningtable	72.26	71.63^-0.63	73.96^+1.70
Sofa	70.44	75.12^+4.68	73.31^+2.87
Boat	65.69	62.92^-2.77	66.60^+0.91
Chair	54.19	51.53^-2.66	56.31^+2.12
Bottle	52.06	57.03^+4.97	56.48^+4.42
Pottedplant	45.96	43.85^-2.11	45.25^-0.71

Table 4. Comparison of detection accuracy of 20 target classes

Category	FR	FR+MA1	FR+MA2
mAP	74.25	74.61^+0.36	74.70^+0.45
Cat	87.23	87.10^-0.13	90.07^+2.84
Car	86.60	85.94^-0.66	84.66^-1.94
Horse	84.87	85.20^+0.33	85.90^+1.03
Dog	83.62	86.41^+2.79	85.11^+1.49
Bus	80.29	83.17^+2.88	85.36^+5.07
Train	82.49	81.98^-0.51	82.09^-0.40
Motorbike	83.48	81.38^-2.10	81.37^-2.11
Bicycle	79.56	81.80^+2.24	83.38^+3.82
Person	79.75	78.64^-1.11	79.61^-0.14
Aeroplane	79.75	77.68^-2.07	76.20^-3.55
Sheep	75.22	75.62^+0.40	73.97^-1.25
Bird	74.14	75.58^+1.44	75.87^+1.73
Cow	74.73	78.09^+3.36	73.06^-1.67
Tvmonitor	74.34	72.20^-2.14	73.79^-0.55
Diningtable	72.26	70.78^-1.48	70.58^-1.68
Sofa	70.44	74.22^+3.78	73.24^+2.80
Boat	65.69	66.58^+0.89	66.29^+0.66
Chair	54.19	53.49^-0.70	51.81^-2.38
Bottle	52.06	51.17^-0.89	55.37^+3.31
Pottedplant	45.96	45.20^-0.76	46.21^+0.25

Table 5. Comparison of experimental results of ablation of MA

Qisheng Wang, Fengsui Wang, Jingang Chen, Furong Liu. Faster R-CNN Target-Detection Algorithm Fused with Adaptive Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215016

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information