基于改进DeepLabV3+的煤尘图像分割方法

左纯子; 王征; 张科; 潘红光

doi:10.13272/j.issn.1671-251x.2021120086

基于改进DeepLabV3+的煤尘图像分割方法

doi: 10.13272/j.issn.1671-251x.2021120086

西安科技大学电气与控制工程学院, 陕西西安　710054

基金项目: 国家自然科学基金资助项目(51804249)。

详细信息

作者简介:
左纯子(1996—)，女，陕西西安人，硕士研究生，研究方向为图像处理，E-mail：1220630049@qq.com

中图分类号: TD67
计量
- 文章访问数: 275
- HTML全文浏览量: 125
- PDF下载量: 40
- 被引次数: 0
出版历程
- 收稿日期: 2021-12-27
- 修回日期: 2022-04-25
- 网络出版日期: 2022-03-05

Coal dust image segmentation method based on improved DeepLabV3+

College of Electrical and Control Engineering, Xi'an University of Science and Technology, Xi'an 710054, China

摘要

摘要: 采用传统的语义分割网络对煤尘颗粒这种较小的目标进行分割时存在深层信息易丢失、细节提取不明显等问题。针对该问题，提出了一种基于改进DeepLabV3+的煤尘图像分割方法。从3个方面对DeepLabV3+网络模型进行改进：① 在编码器中，用CA−MobileNetV3轻量化模块代替Xception实现特征提取，确保特征提取更加细致、准确。② 在空洞空间卷积池化金字塔(ASPP)模块中对空洞率进行改进，使其更适合小颗粒煤尘提取。③ 在解码器中引入全局注意力上采样(GAU)模块，在计算量较小时对低层特征信息进行加权，用高层特征信息指导低层特征信息，实现特征融合。GAU模块用全局上采样机制代替解码器的上采样机制，使煤尘颗粒的特征信息经过长距离传输后不衰减，更加有利于捕捉煤尘颗粒的边缘细节信息。实验结果表明，改进DeepLabV3+网络模型在煤尘数据集上的召回率为90.26%，准确度为89.23%，相比于其他网络模型，改进DeepLabV3+对煤尘特征的学习能力更强，能获取更多细节信息，并大幅缩短训练时间，对小目标的分割效果更优。
- 煤尘 /
- 图像分割 /
- 特征提取 /
- DeepLabV3+ /
- CA−MobileNetV3 /
- 空洞卷积 /
- 金字塔结构 /
- 全局注意力上采样
Abstract: When the traditional semantic segmentation network is used to segment the small coal dust particles, there are some problems such as easy loss of deep information and unclear detail extraction. In order to solve this problem, a coal dust image segmentation method based on improved DeepLabV3+ is proposed. DeepLabV3+ network model is improved in three aspects. ① In the encoder, the CA-MobileNetV3 lightweight module is used to replace Xception to achieve characteristic extraction and ensure more detailed and accurate characteristic extraction. ② The atrous rate is improved in the atrous spatial pyramid pooling(ASPP) module to make it more suitable for extracting small coal dust particles. ③ A global attention up-sample(GAU) module is introduced into the decoder to weight the low-level characteristic information when the calculation amount is small. And the high-level characteristic information is used to guide the low-level characteristic information to realize characteristic fusion. The GAU module uses a global up-sampling mechanism to replace the up-sampling mechanism of the decoder. The characteristic information of the coal dust particles is not attenuated after long-distance transmission. And the method is more conducive to capture the edge detail information of the coal dust particles. The experimental results show that the recall rate of the improved DeepLabV3+ network model on the coal dust data set is 90.26%, and the accuracy is 89.23%. Compared with other network models, the improved DeepLabV3+ network model can effectively enhance the learning ability of coal dust characteristics, obtain more detailed information, greatly shorten the training time, and has better segmentation effect on small targets.
- coal dust /
- image segmentation /
- characteristic extraction /
- DeepLabV3+ /
- CA−MobileNetV3 /
- atrous convolution /
- pyramid structure /
- global attention up-sample

HTML全文

图 1 改进DeepLabV3+网络模型结构

Figure 1. The structure of improved DeepLabV3+ network model

下载: 全尺寸图片幻灯片

图 2 CA−MobileNetV3模块结构

Figure 2. The structure of CA-MobileNetV3 module

下载: 全尺寸图片幻灯片

图 3 GAU模块结构

Figure 3. The structure of GAU module

下载: 全尺寸图片幻灯片

图 4 图像特征提取结果

Figure 4. Image characteristic extraction results

下载: 全尺寸图片幻灯片

图 5 不同网络模型对煤尘图像的分割效果

Figure 5. Segmentation effect of different network models on coal dust image

下载: 全尺寸图片幻灯片

表 1 不同特征提取网络的性能

Table 1. Performance of different characteristic extraction networks

特征提取网络	MIoU/%	运行时间/h	模型大小/MB
CA−MobileNetV3	72.36	1.28	507.18
Xception	78.43	1.01	320.87

下载: 导出CSV

表 2 不同空洞率下DeepLabV3+网络模型的分割性能

Table 2. Segmentation performance of DeepLabV3+ network model under different dilation rates

空洞率	PA/%	MIoU/%
[1，6，8，12]	84.23	56.53
[1，12，18，24]	84.53	56.23
[1，5，7，11]	86.63	60.03
[1，7，11，13]	85.26	58.73
[1，3，7，9]	85.36	58.63

下载: 导出CSV

表 3 GAU模块性能

Table 3. The performance of GAU module

网络模型	MIoU/%	准确度/%
未引入GAU模块	72.36	84.13
引入GAU模块	76.56	85.27

下载: 导出CSV

表 4 各网络模型性能指标

Table 4. Performance indicators of each network model

网络模型	召回率%	准确度%	F₁/%	MIoU/%	耗时/h	占用内存/GB
U−Net	85.34	82.32	83.80	81	1.50	8.7
Unet−SE	87.21	83.29	85.23	84	1.32	8.6
SegNet	86.23	79.03	82.47	83	1.47	8.9
PSPNet	85.78	83.16	84.45	89	1.35	8.1
FCN	84.45	78.96	81.61	82	1.41	7.7
DeepLabV3+	86.67	84.13	85.38	90	1.25	8.0
改进 DeepLabV3+	90.26	89.23	89.74	93	1.02	7.5

下载: 导出CSV

参考文献(17)

[1]	王国法,赵国瑞,任怀伟. 智慧煤矿与智能化开采关键核心技术分析[J]. 煤炭学报,2019,44(1):34-41. WANG Guofa,ZHAO Guorui,REN Huaiwei. Analysis on key technologies of intelligent coal mine and intelligent mining[J]. Journal of China Coal Society,2019,44(1):34-41.
[2]	韩建国. 神华智能矿山建设关键技术研发与示范[J]. 煤炭学报,2016,41(12):3181-3189. HAN Jianguo. Key technology research and demonstration of intelligent mines in Shenhua Group[J]. Journal of China Coal Society,2016,41(12):3181-3189.
[3]	PLESSIS J J L D. Active explosion barrier performance against methane and coal dust explosions[J]. International Journal of Coal Science and Technology,2015(4):261-268.
[4]	张锦仁. 选煤技术的现状及发展趋势探索[J]. 内蒙古煤炭经济,2020(6):194. doi: 10.3969/j.issn.1008-0155.2020.06.133 ZHANG Jinren. Exploration on the status quo and development trend of coal preparation technology[J]. Inner Mongolia Coal Economy,2020(6):194. doi: 10.3969/j.issn.1008-0155.2020.06.133
[5]	吴开兴,宋剑. 基于灰度共生矩阵的煤与矸石自动识别研究[J]. 煤炭工程,2016,48(2):98-101. WU Kaixing,SONG Jian. Automatic coal-gangue identification based on gray level co-occurrence matrix[J]. Coal Engineering,2016,48(2):98-101.
[6]	郜亚松,张步勤,郎利影. 基于深度学习的煤矸石识别技术与实现[J]. 煤炭科学技术,2021,49(12):202-208. GAO Yasong,ZHANG Buqin,LANG Liying. Coal and gangue recognition technology and implementation based on deep learning[J]. Coal Science and Technology,2021,49(12):202-208.
[7]	ZHOU Hao, ZHANG Jun, LEI Jun, et al. Image semantic segmentation based on FCN-CRF model[C]//International Conference on Image, Vision and Computing, Portsmouth, 2016: 9-14
[8]	ZHAO Hengshuang, SHI Jianping, QI Xiaojuan, et al. Pyramid scene parsing network[C]// IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, 2016.
[9]	BADRINARAYANAN V,KENDALL A,CIPOLLA R. SegNet:a deep convolutional encoder-decoder architecture for image segmentation[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence,2017,39(12):2481-2495.
[10]	RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[M]//NAVAB N, HORNEGGER J, WELLS W M, et al. Medical image computing and computer-assisted intervention, Springer, Cham, 2015: 234-241.
[11]	王征,张赫林,李冬艳. 特征压缩激活作用下U−Net网络的煤尘颗粒特征提取[J]. 煤炭学报,2021,46(9):3056-3065. WANG Zheng,ZHANG Helin,LI Dongyan. Feature extraction of coal dust particles based on U-Net combined with squeeze and excitation module[J]. Journal of China Coal Society,2021,46(9):3056-3065.
[12]	CHEN L C, ZHU Y, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[EB/OL]. （2018-08-22） [2021-11-20]. https://arxiv.org/abs/1802.02611.
[13]	HOU Qibin, ZHOU Daquan, FENG Jiashi. Coordinate attention for efficient mobile network design[EB/OL]. （2021-03-04） [2021-11-20]. https://arxiv.org/abs/2103.02907.
[14]	CHEN L C,PAPANDREOU G,KOKKINOS I,et al. DeepLab:semantic image segmentation with deep convolutional nets,atrous convolution,and fully connected CRFs[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2018,40(4):834-848. doi: 10.1109/TPAMI.2017.2699184
[15]	翁和王,叶球孙. 图像处理中特征提取的应用及增强算法研究[J]. 重庆理工大学学报(自然科学),2016,30(7):119-122. WENG Hewang,YE Qiusun. Applications of feature extraction and enhancement algorithm in image processing[J]. Journal of Chongqing University of Technology(Natural Science),2016,30(7):119-122.
[16]	常雪昕,韩军,廖子豪. 基于语义分割的输电线路螺丝识别的研究与实现[J]. 工业控制计算机,2019,32(8):118-120. doi: 10.3969/j.issn.1001-182X.2019.08.046 CHANG Xuexin,HAN Jun,LIAO Zihao. Transmission line screw recognition based on semantic segmentation[J]. Industrial Control Computer,2019,32(8):118-120. doi: 10.3969/j.issn.1001-182X.2019.08.046
[17]	屈航, 嵇启春, 段中兴. 改进Deeplab V3+网络在视觉SLAM三维地图构建应用[J/OL]. 小型微型计算机系统: 1-6[2021-12-07]. http://kns.cnki.net/kcms/detail/21.1106.TP.20210818.1048.026.html. QU Hang, JI Qichun, DUAN Zhongxing. Improved Deeplab V3+ network application for visual SLAM 3D map construction[J/OL]. Journal of Chinese Computer Systems: 1-6[2021-12-07]. http://kns.cnki.net/kcms/detail/21.1106.TP.20210818.1048.026.html.