Bullet Appearance Defect Detection Based on Improved Faster Region-Convolutional Neural Network

Xiaoyun Ma; Dan Zhu; Chen Jin; Xinxin Tong

doi:10.3788/LOP56.151202

Laser & Optoelectronics Progress, Volume. 56, Issue 15, 151202(2019)

Bullet Appearance Defect Detection Based on Improved Faster Region-Convolutional Neural Network

Xiaoyun Ma^{1,2,3,4,5、*}, Dan Zhu^1,2,3,4,5, Chen Jin^1,2,4,5, and Xinxin Tong^1,2,4,5

Author Affiliations

¹ Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China

² Institutes for Robotics and Intelligent Manufacturing, Chinese Academy of Sciences, Shenyang, Liaoning 110016, China

³ University of Chinese Academy of Sciences, Beijing 100049, China

⁴ Key Laboratory of Opto-Electronic Information Processing, CAS, Shenyang, Liaoning 110016, China

⁵ The Key Lab of Image Understanding and Computer Vision, Shenyang, Liaoning 110016, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(11)

Fig. 1. Target detection framework based on Faster R-CNN

Download full size

Fig. 2. Structure of region proposal network

Download full size

Fig. 3. RoI pooling process

Download full size

Fig. 4. Diagram of anchor proportion

Download full size

Fig. 5. K-means++ clustering results based on different g values. (a) g=3; (b) g=4; (c) g=5; (d) g=6; (e) g=7

Download full size

Fig. 6. Examples of bullet appearance defect dataset. (a) Mouthcrack; (b) mouthgap

Download full size

Fig. 7. Partial test results based on improved Faster R-CNN model. (a) Mouthcrack; (b) mouthgap

Download full size

Table 1. Experimental environment
View table
Table 1. Experimental environment
Name Type
System LINUX64 Ubuntu14.04
Frame Caffe
Language Python,C++,Protobuf
CPU Intel Core i7-7700
GPU GTX1080Ti
Memory/GB 11
RAM/GB 16
Hard disk/GB 250

Table 2. Comparison of detection results of Faster R-CNN model based on different convolutional networks
View table
Table 2. Comparison of detection results of Faster R-CNN model based on different convolutional networks
Method mAP /%
Faster R-CNN+ZFNet 91.09
Fater R-CNN+ VGG_CNN_M_1024 93.82
Fater R-CNN+ VGG16 96.37

Table 3. Comparison of test results of different anchor generation methods
View table
Table 3. Comparison of test results of different anchor generation methods
Method Parameter mAP /%
Anchor boxes k=9 96.37
K-means++ g=3 90.03
K-means++ g=4 94.68
K-means++ g=5 97.24
K-means++ g=6 98.06
K-means++ g=7 98.02

Table 4. Comparison of test results of different anchor numbers
View table
Table 4. Comparison of test results of different anchor numbers
Method Parameter Proposals mAP /% Speed /(frame·s^-1)
Anchor boxes k=9 300 96.37 17
K-means++ g=6 300 98.06 19
K-means++ g=6 100 97.75 28

Tools

Get Citation

Copy Citation Text

Xiaoyun Ma, Dan Zhu, Chen Jin, Xinxin Tong. Bullet Appearance Defect Detection Based on Improved Faster Region-Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2019, 56(15): 151202

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites