Classification of Ultrasonic Infrared Thermal Images Using a Convolutional Neural Network
-
摘要: 在超声红外热像技术应用中,从红外热图像来判断被测对象是否含有裂纹,通常需要先基于人工经验,从红外热图像中提取特征再采用某种模式识别方法进行分类,裂纹的识别与定位过程繁琐且识别率较低。为此,提出一种基于卷积神经网络技术的超声红外热图像裂纹检测与识别方法,其特点是可以直接从超声红外图像中学习特征进而实现是否含有裂纹红外热图像的分类。通过实验得到的含裂纹和不含裂纹金属平板试件的红外热图像,建立卷积神经网络模型对图像中是否含有裂纹进行分类,研究结果表明,参数优化后的卷积神经网络模型对超声红外热图像的有无裂纹分类准确率达到98.7%。Abstract: In the application of ultrasonic infrared thermographic technology, it is usually necessary to extract features from infrared thermographic images based on artificial experience and then adopt a pattern recognition method to classify the cracks. The identification and positioning process of the cracks is complicated, and the recognition rate is low. Therefore, a method of crack detection and recognition in ultrasonic infrared thermal images based on convolutional neural network technology is proposed in this paper. Its feature is that the features can be directly learned from the ultrasonic infrared image to realize the classification of infrared thermal images containing cracks. Thesis through the research experiment of metal plate specimen of the crack in and do not contain infrared thermal images, the convolutional neural network model is established for whether the image contains crack classification, the results show that the parameter optimized convolution neural network model for ultrasonic infrared thermal images of crack classification accuracy rate reached 98.7%.
-
表 1 网络模型参数总量
Table 1. Total parameters of network model
Layer’s name Explanation Number of parameters Input Input - Cov1 96 convolution kernels(11×11×3) 11×11×3×96+96=34944 Cov2 256 convolution kernels(5×5×48) (5×5×48×128+128)×2=307456 Cov3 384 convolution kernels(3×3×256) 3×3×256×384+384=885120 Cov4 384 convolution kernels(3×3×192) (3×3×192×192+192)×2=663936 Cov5 256 convolution kernels(3×3×192) (3×3×192×128+128)×2=442624 Fc1 Full connection layer 1 (6×6×128×2)×4096+4096=37752832 Fc2 Full connection layer 2 4096×4096+4096=16781312 Output Softmax classification output 4096×1000+1000=4097000 Summation - 60965224 表 2 不同批量尺寸下的识别正确率和网络训练时间
Table 2. Recognition accuracy and network training time underdifferent batch sizes
Batch size 16 32 64 128 Accuracy/% 56.69 97.34 97.64 97.11 Time 7 min43 s 8 min10 s 9 min15 s 9 min18 s 表 3 批量尺寸32时不同丢失比率下的识别率
Table 3. Different dropout rate recognition rates at batch size 32
Dropout 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Accuracy/% 95.6 97.4 97.6 97.5 96.9 98.4 97.3 98.2 97.1 97.2 表 4 批量尺寸64时不同丢失比率下的识别率
Table 4. Different dropout rate recognition rates at batch size 64
Dropout 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Accuracy/% 96.6 98.1 98.3 97.1 97.3 98.7 97.1 98.1 97.4 97.3 表 5 不同尺寸图像识别率
Table 5. Image recognition rates of different sizes
The size of the image 16×16 32×32 64×64 128×128 Accuracy/% 98.38 98.36 98.70 98.03 Time 11 min12 s 8 min2 s 9 min18 s 13 min29 s 表 6 各神经网络模型对比
Table 6. Comparison of each neural network model
Neural Network Model Accuracy/% Time AlexNet 98.70 9min18s GoogLeNet 86.67 10min28s ResNet 96.43 9min04s -
[1] 曾平平, 李林升. 基于卷积神经网络的水果图像分类识别研究[J]. 机械设计与研究, 2019, 35(1): 23-26, 34. https://www.cnki.com.cn/Article/CJFDTOTAL-JSYY201901010.htmZENG Pingping, LI Linsheng. Classification and Recognition of Common Fruit Images Based on Convolutional Neural Network[J]. Machine Design & Research, 2019, 35(1): 23-26, 34. https://www.cnki.com.cn/Article/CJFDTOTAL-JSYY201901010.htm [2] 林明旺. 基于卷积神经网络的鱼类图像识别与分类[J]. 电子技术与软件工程, 2017(6): 82-83. https://www.cnki.com.cn/Article/CJFDTOTAL-DZRU201706065.htmLIN Mingwang. Fish image recognition and classification based on convolutional neural network[J]. Electronic Technology & Software Engineering, 2017(6): 82-83. https://www.cnki.com.cn/Article/CJFDTOTAL-DZRU201706065.htm [3] 张安安, 黄晋英, 冀树伟, 等. 基于卷积神经网络图像分类的轴承故障模式识别[J]. 振动与冲击, 2020, 39(4): 165-171. https://www.cnki.com.cn/Article/CJFDTOTAL-ZDCJ202004021.htmZHANG An'an, HUANG Jinying, JI Shuwei, et al. Bearing fault pattern recognition based on image classification with CNN[J]. Journal of Vibration and Shock, 2020, 39(4): 165-171. https://www.cnki.com.cn/Article/CJFDTOTAL-ZDCJ202004021.htm [4] 李玉鑑, 张婷, 单传辉, 等. 深度学习卷积神经网络从入门到精通[M]. 北京: 机械工业出版社, 2018.LI Yujian, ZHANG Ting, SHAN Chuanhui, et al. Deep Learning Convolutional Neural Network From Entry to Mastery[M]. Beijing: China Machine Press, 2018. [5] 李彦冬, 郝宗波, 雷航. 卷积神经网络研究综述[J]. 计算机应用, 2016, 36(9): 2508-2515, 2565. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY201609029.htmLI Yandong, HAO Zongbo, LEI Hang. Survey of convolutional neural network[J]. Journal of Computer Applications, 2016, 36(9): 2508-2515, 2565. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJY201609029.htm [6] 冯辅周, 张超省, 宋爱斌, 等. 超声红外热像检测中疲劳裂纹的检出概率模型研究[J]. 红外与激光工程, 2016, 45(3): 60-65. https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201603008.htmFENG Fuzhou, ZHANG Chaosheng, SONG Aibin, et al. Probability of detection model for fatigue crack in ultrasonic infrared imaging[J]. Infrared and Laser Engineering, 2016, 45(3): 60-65. https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201603008.htm [7] 冯辅周, 张超省, 闵庆旭, 等. 超声红外热像技术中金属平板裂纹的生热特性[J]. 红外与激光工程, 2015, 44(5): 1456-14461. https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201505012.htmFENG Fuzhou, ZHANG Chaosheng, MIN Qingxu, et al. Heating characteristics of metal plate crack in sonic IR imaging[J]. Infrared and Laser Engineering, 2015, 44(5): 1456-14461. https://www.cnki.com.cn/Article/CJFDTOTAL-HWYJ201505012.htm [8] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with deep convolutional neural networks[C]//International Conference on Neural Information Processing Systems, 2012: 1106-1114. [9] Szegedy C, LIU W, JIA Y, et al. Going deeper with convolutions[C]//Proceedings of the 2015 IEEE Conference on Computer Vision and Pattern Recognition. Washington, DC: IEEE Computer Society, 2015: 1-8. [10] HE K, ZHANG X, REN S, et al. Deep Residual Learning for Image Recognition[EB/OL]. [2020-6-20]. https://arxiv.org/pdf/1512.03385.pdf.