Detection Method of Downpipe Diseases Based on Visual Attention Mechanism

Jiasong Zhu; Tianzhu Ma; Haokun Yang; Xu Fang; Qing Li

doi:10.3788/LOP202259.1815001

Laser & Optoelectronics Progress, Volume. 59, Issue 18, 1815001(2022)

Detection Method of Downpipe Diseases Based on Visual Attention Mechanism

Jiasong Zhu^1,2、**, Tianzhu Ma^1,3、***, Haokun Yang^1,3, Xu Fang^2,4, and Qing Li^1、*

Author Affiliations

¹Institute of Urban Smart Transportation & Safety Maintenance, Shenzhen 518000, Guangdong , China

²Shenzhen Institute of Artificial Intelligence and Robotics for Society, Shenzhen 518000, Guangdong , China

³College of Civil and Transportation Engineering, Shenzhen University, Shenzhen 518000, Guangdong , China

⁴College of Electronics and Information Engineering, Shenzhen University, Shenzhen 518000, Guangdong , China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(7)

Fig. 1. Residual attention model network diagram

Download full size

Fig. 2. Schematic diagram of fusion strategy

Download full size

Table 1. Results of different defects under shallow fusion strategy
View table
Table 1. Results of different defects under shallow fusion strategy
Style Precision Recall F1 Number of samples
Broken 0.82 0.78 0.80 400
Corrosion 0.89 0.90 0.90 400
Deformation 0.88 0.86 0.87 400
Normal 0.92 0.97 0.95 400

Table 2. Results of different defects under middle fusion strategy
View table
Table 2. Results of different defects under middle fusion strategy
Style Precision Recall F1 Number of samples
Broken 0.82 0.73 0.77 400
Corrosion 0.88 0.94 0.91 400
Deformation 0.84 0.80 0.82 400
Normal 0.89 0.97 0.93 400

Table 3. Results of different defects under bottom fusion strategy
View table
Table 3. Results of different defects under bottom fusion strategy
Style Precision Recall F1 Number of samples
Broken 0.86 0.86 0.86 400
Corrosion 0.88 0.91 0.89 400
Deformation 0.87 0.83 0.85 400
Normal 0.98 0.99 0.98 400

Table 4. Results of different fusion strategies
View table
Table 4. Results of different fusion strategies
Neural network fusion strategy F1 accuracy recall
Residual Shallow fusion 0.860 0.859 0.859
Attention Middle fusion 0.878 0.879 0.878
Network Bottom fusion 0.895 0.896 0.896

Table 5. Results of different neural networks based on bottom fusion strategy
View table
Table 5. Results of different neural networks based on bottom fusion strategy
Neural network F1 Accuracy Recall
Residual Attention Network 0.895 0.896 0.896
Resnet50 0.874 0.876 0.875

Tools

Get Citation

Copy Citation Text

Jiasong Zhu, Tianzhu Ma, Haokun Yang, Xu Fang, Qing Li. Detection Method of Downpipe Diseases Based on Visual Attention Mechanism[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites