煤体红外热像异常区域分割方法

赵小虎; 车亭雨; 叶圣; 田贺; 张凯

doi:10.13272/j.issn.1671-251x.2022030086

煤体红外热像异常区域分割方法

doi: 10.13272/j.issn.1671-251x.2022030086

赵小虎^{1, 2,},
车亭雨^{1, 2},
叶圣^{1, 2},
田贺³,
张凯^{1, 2, ,}

1.
矿山互联网应用技术国家地方联合工程实验室, 江苏徐州　221008
2.
中国矿业大学信息与控制工程学院, 江苏徐州　221008
3.
中国矿业大学安全工程学院, 江苏徐州　221008

基金项目: 中央高校基本科研业务费专项资金资助项目(2020ZDPY0223)。

详细信息

作者简介:
赵小虎（1976—），男，江苏徐州人，教授，博士，主要研究方向为矿山物联网、计算机网络和智能计算，E-mail：xiaohuzhao@126.com

通讯作者:
张凯（1982—），男，江苏徐州人，博士研究生，研究方向为智慧矿山网络优化，E-mail：kaizhang@cumt.edu.cn

中图分类号: TD315/67
计量
- 文章访问数: 263
- HTML全文浏览量: 50
- PDF下载量: 32
- 被引次数: 0
出版历程
- 收稿日期: 2022-03-28
- 修回日期: 2022-09-06
- 网络出版日期: 2022-06-21

Segmentation method of the abnormal area of coal infrared thermal image

ZHAO Xiaohu^{1, 2
,},
CHE Tingyu^{1, 2},
YE Sheng^{1, 2},
TIAN He³,
ZHANG KAI^{1, 2
, ,}

1.
The National Joint Engineering Laboratory of Internet Applied Technology of Mines, Xuzhou 221008, China
2.
School of Information and Control Engineering, China University of Mining and Technology, Xuzhou 221008, China
3.
School of Safety Engineering, China University of Mining and Technology, Xuzhou, 221008, China

摘要

摘要: 红外辐射可反映煤岩受载破坏情况，用于监测和预防煤岩动力灾害，但红外热像仪生成的红外热像图像素分辨率低、噪声较大，导致检测结果受主观因素影响较大，无法准确识别煤体损伤区域。将深度学习和红外热像结合进行无损检测已成为趋势，但目前结合深度学习和红外热像对煤体受载破坏进行识别检测的研究相对较少。针对上述问题，提出一种基于多尺度通道注意力模块(MS−CAM)U−Net模型的煤体红外热像异常区域分割方法。在传统U−Net模型的编码器中引入MS−CAM，设计了基于MS−CAM的U−Net模型结构，使模型在关注煤体红外热像异常区域显著特征的同时，还关注异常区域小目标特征，以提高异常区域分割精度。为降低煤体红外热像数据集匮乏对模型准确率和适用性的影响，对创建的煤体红外热像数据集进行数据增强操作，并采用MS COCO数据集对基于MS−CAM的U−Net模型进行预训练，再采用煤体红外热像数据集训练，得出最终网络权重。实验结果表明，该方法可有效分割煤体红外热像异常区域，精确率、F1分数、Dice系数和平均交并比分别为94.75%，94.94%，94.65%，90.03%，均优于Deeplab模型、U−Net模型和基于SENet注意力机制的U−Net模型。
- 煤岩动力灾害 /
- 煤岩受载破坏 /
- 红外辐射 /
- 红外热像 /
- 异常区域分割 /
- U−Net模型 /
- 多尺度通道注意力模块 /
- 深度学习
Abstract: Infrared radiation can reflect the damage of coal and rock under load, and can be used to monitor and prevent the dynamic disaster of coal and rock. But the infrared thermal image generated by the infrared thermal imager has low pixel resolution and large noise, which leads to the detection result being greatly affected by subjective factors. Therefore, the damaged area of the coal body cannot be accurately identified. It has become a trend to combine deep learning with infrared thermal imaging for nondestructive testing. But the research on the identification and detection of coal damage under load by combining deep learning and infrared thermal imaging is relatively few. In order to solve the above problems, a segmentation method of the abnormal area of coal infrared thermal image based on multi-scale channel attention module (MS-CAM) U-Net model is proposed. The MS-CAM is introduced into the encoder of the traditional U-Net model, and the U-Net model structure based on MS-CAM is designed. The model not only pays attention to the major characteristics of the coal infrared thermal image abnormal area, but also pays attention to the small target characteristics of the abnormal area, so as to improve the segmentation accuracy of the abnormal area. In order to reduce the influence of the lack of coal infrared thermal image data set on the accuracy and applicability of the model, the data enhancement operation is carried out on the created coal infrared thermal image data set. The MS-CAM-based U-Net model is pre-trained by using the MS COCO data set. Then the coal infrared thermal image data set is used for training to obtain the final network weight. The experimental result shows that the method can effectively segment the abnormal areas of the infrared thermal image of the coal body. The accuracy rate, the F1 score, the Dice coefficient and the average cross-combination ratio are 94.75%, 94.94%, 94.65%, and 90. 03% respectively. The results are superior to the Deeplab model, the U-Net model and the U-Net model based on the attention mechanism of the SENet.
- dynamic disaster of coal and rock /
- damage of coal and rock under load /
- infrared radiation /
- infrared thermal imaging /
- abnormal area segmentation /
- U-Net model /
- multi-scale channel attention module /
- deep learning

HTML全文

图 1 基于MS−CAM的U−Net模型结构

Figure 1. U-Net model structure based on multi-scale channel attention module(MS-CAM)

下载: 全尺寸图片幻灯片

图 2 MS-CAM结构

Figure 2. MS-CAM structure

下载: 全尺寸图片幻灯片

图 3 实验系统

Figure 3. Experimental system

下载: 全尺寸图片幻灯片

图 4 煤样3种加载破坏时期红外热像图

Figure 4. Three infrared thermal images of coal samples during loading pressure period

下载: 全尺寸图片幻灯片

图 5 LabelMe工具中煤样红外热像异常区域标注

Figure 5. Abnormal area tagging in infrared thermal images of coal samples in LabelMe tool

下载: 全尺寸图片幻灯片

图 6 部分增强图像

Figure 6. Partial enhanced images

下载: 全尺寸图片幻灯片

图 7 损失函数曲线

Figure 7. Loss function curve

下载: 全尺寸图片幻灯片

图 8 实验流程

Figure 8. Experiment process

下载: 全尺寸图片幻灯片

图 9 不同模型对煤体红外热像异常区域的分割结果

Figure 9. Segmentation results of infrared thermal images of coal samples by different models

下载: 全尺寸图片幻灯片

图 10 U−Net（SENet）模型和U−Net（MS−CAM）模型分割结果的类激活热力图

Figure 10. Class activation heat map of segmentation results by U-Net（SENet） model and U-Net（MS-CAM）model

下载: 全尺寸图片幻灯片

表 1 基于MS−CAM的U−Net模型网络结构及对应特征图

Table 1. Network structures of U-Net model based on MS-CAM and corresponding characteristic images

网络结构	特征图尺寸	卷积核参数
Conv_1	256×256×64 256×256×64	3×3×64 3×3×64
MS−CAM_1	256×256×64	−
DownSampling_1	128×128×64	2×2
Conv_2	128×128×128 128×128×128	3×3×128 3×3×128
MS−CAM_2	128×128×128	−
DownSampling_2	64×64×128	2×2
Conv_3	64×64×256 64×64×256	3×3×256 3×3×256
MS−CAM_3	64×64×256	−
DownSampling_3	32×32×256	2×2
Conv_4	32×32×512 32×32×512	3×3×512 3×3×512
MS−CAM_4	32×32×512	−
DownSampling_4	16×16×512	2×2
Conv	16×16×512 16×16×512	3×3×512 3×3×512
UpSampling_1	32×32×512	2×2
Concatenate_1	32×32×1024	UpSampling_1+ MS−CAM_4
conv_1	32×32×256 32×32×256	3×3×256 3×3×256
UpSampling_2	64×64×256	2×2
Concatenate_2	64×64×512	UpSampling_2+ MS−CAM_3
conv_2	64×64×128 64×64×128	3×3×128 3×3×128
UpSampling_3	128×128×128	2×2
Concatenate_3	128×128×256	UpSampling_3+ MS−CAM_2
conv_3	128×128×64 128×128×64	3×3×64 3×3×64
UpSampling_4	256×256×64	2×2
Concatenate_4	256×256×128	UpSampling_4+ MS−CAM_1
conv_4	256×256×64 256×256×64	3×3×64 3×3×64
Conv1×1	256×256×1	1×1×1

下载: 导出CSV

表 2 不同模型分割结果评价指标对比

Table 2. Comparison of evaluation indexes for segmentation results of different models %

模型	精确率	F1分数	Dice系数	MIoU
Deeplab	88.65	90.31	87.92	83.78
U−Net	92.28	91.67	91.58	84.85
U−Net(SENet)	93.46	93.56	92.81	87.28
U−Net（MS−CAM）	94.75	94.94	94.65	90.03

下载: 导出CSV

参考文献(20)

[1]	程富起,李忠辉,魏洋,等. 基于单轴压缩红外辐射的煤岩损伤演化特征[J]. 工矿自动化,2018,44(5):64-70. doi: 10.13272/j.issn.1671-251x.2017110064 CHENG Fuqi,LI Zhonghui,WEI Yang,et al. Coal-rock damage evolution characteristics based on infrared radiation under uniaxial compression[J]. Industry and Mine Automation,2018,44(5):64-70. doi: 10.13272/j.issn.1671-251x.2017110064
[2]	娄全,李忠辉,李爱国,等. 混凝土变形破坏的红外辐射特征研究[J]. 工矿自动化,2015,41(7):44-48. LOU Quan,LI Zhonghui,LI Aiguo,et al. Research of infrared radiation characteristics of concrete deformation and failure[J]. Industry and Mine Automation,2015,41(7):44-48.
[3]	TIAN He,LI Zhonghui,SHEN Xiaofan,et al. Identification method of infrared radiation precursor information of coal sample failure and instability under uniaxial compression[J]. Infrared Physics & Technology,2021,119:103957.
[4]	SUN Hai,MA Liqiang,LIU Wei,et al. The response mechanism of acoustic and thermal effect when stress causes rock damage[J]. Applied Acoustics,2021,180:108093. doi: 10.1016/j.apacoust.2021.108093
[5]	LIU Wei,MA Liqiang,SUN Hai,et al. Using the characteristics of infrared radiation b-value during the rock fracture process to offer a precursor for serious failure[J]. Infrared Physics & Technology,2021,114:103644.
[6]	宋晶晶,李忠辉,张昕,等. 岩样损伤红外热像的归一化直方图表征研究[J]. 红外技术,2021,43(8):777-783. SONG Jingjing,LI Zhonghui,ZHANG Xin,et al. Research on normalized histogram characterization of infrared thermal image of rock sample damage[J]. Infrared Technology,2021,43(8):777-783.
[7]	CAO K,YUAN Q,XIE G,et al. Infrared radiation characteristics during crack development in water-bearing sandstone[J]. Soil Mechanics and Foundation Engineering,2021,58(3):209-214. doi: 10.1007/s11204-021-09730-2
[8]	LI Zhonghui,YIN Shan,NIU Yue,et al. Experimental study on the infrared thermal imaging of a coal fracture under the coupled effects of stress and gas[J]. Journal of Natural Gas Science and Engineering,2018,55:444-451. doi: 10.1016/j.jngse.2018.05.019
[9]	ZHANG Xueliang,WANG Deliang. Deep learning based binaural speech separation in reverberant environments[J]. IEEE/ACM Transactions on Audio,Speech,and Language Processing,2017,25(5):1075-1084. doi: 10.1109/TASLP.2017.2687104
[10]	WANG Bin,DONG Ming,REN Ming,et al. Automatic fault diagnosis of infrared insulator images based on image instance segmentation and temperature analysis[J]. IEEE Transactions on Instrumentation and Measurement,2020,69(8):5345-5355. doi: 10.1109/TIM.2020.2965635
[11]	AKRAM M W,LI Guiqiang,JIN Yi,et al. Automatic detection of photovoltaic module defects in infrared images with isolated and develop-model transfer deep learning[J]. Solar Energy,2020,198:175-186. doi: 10.1016/j.solener.2020.01.055
[12]	BANG H T,PARK S,JEON H. Defect identification in composite materials via thermography and deep learning techniques[J]. Composite Structures,2020,246:112405. doi: 10.1016/j.compstruct.2020.112405
[13]	DAI Yimian, GIESEKE F, OEHMCKE S, et al. Attentional feature fusion[C]. The IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, 2021: 3560-3569.
[14]	IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]. International Conference on Machine Learning, Lille, 2015: 448-456.
[15]	SHARMA G, LIU D, MAJI S, et al. Parsenet: a parametric surface fitting network for 3D point clouds[C]. European Conference on Computer Vision, Glasgow, 2020: 261-276.
[16]	WU Yuxin,HE Kaiming. Group normalization[J]. International Journal of Computer Vision,2018,128(3):742-755.
[17]	TAN Chuanqi, SUN Fuchuan, KONG Tao, et al. A survey on deep transfer learning[C]. International Conference on Artificial Neural Networks, Rhodes, 2018: 270-279.
[18]	LIN T Y, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]. European Conference on Computer Vision, Zurich, 2014: 740-755.
[19]	MILLETARI F, NAVAB N, AHMADI S A. V-net: fully convolutional neural networks for volumetric medical image segmentation[C]. The 4th International Conference on 3D Vision, Stanford, 2016: 565-571.
[20]	SHELHAMER E,LONG J,DARRELL T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence,2015,39(4):640-651.