一种改进DeeplabV3网络的烟雾分割算法

doi:10.19665/j.issn1001-2400.2019.06.008

摘要/Abstract

摘要：

由于现有的烟雾检测方法大多依靠手工选取特征,往往不能准确地分割出视频图像中的烟雾区域。基于此,提出了改进的DeeplabV3烟雾分割算法。改进的算法在基础编码器网络后添加了特征细化模块来削弱空洞卷积带来的网格效应;针对烟雾这类尺度和姿态多变的非刚性目标,在带有空洞卷积的空间金字塔模块中引入可变形卷积来更好地学习烟雾的形变;为了进一步恢复烟雾的空间细节,提出了通道注意力解码器模块。在烟雾图片数据集的测试下,改进后的模型平均每张图片的预测时间约达到71.73ms,平均像素精确度约达到97.78%,平均交并比约达到91.21%,精度与DeeplabV3模型相比分别提高了0.56%及2.17%,更加适用于烟雾分割。公开的烟雾视频测试结果表明,该模型的检测率高于现有的视频烟雾检测算法,具有一定的实用价值。

关键词: 图像处理, 烟雾检测, 语义分割, 可变形卷积, 注意力机制, 深度学习

Abstract:

Existing smoke detection methods depend mostly on features which are selected manually and the smoke areas in video images often cannot be segmented accurately. This paper proposes an improved DeeplabV3 smoke segmentation algorithm based on this. A feature refinement module is added after the basic encoder network to weaken the gridding effects caused by dilated convolutions. For the non-rigid objects such as smoke with variable scales and postures, the Atrous Spatial Pyramid Pooling module is combined with the deformable convolution to better adapt to the smoke deformation. And a channel attention decoder module is proposed to further restore the spatial details of smoke images. According to the test of the smoke image data set, the proposed model has a faster average prediction speed of 71.73ms per image. Besides, compared with the DeeplabV3 network, this algorithm can lead to a higher MPA (Mean Pixel Accuracy) of 97.78% and a higher MIoU (Mean Intersection over Union) of 91.21%, thus improving the performance by 0.56% and 2.17% respectively, and making it more suitable for smoke segmentation. Public smoke video test results show that this model outperforms other video-based smoke detection methods for the detection rate, and that it is of certain research significance and practical value.

Key words: image processing, smoke detection, semantic segmentation, deformable convolution, attention mechanism, deep learning

中图分类号:

TP391

汪梓艺,苏育挺,刘艳艳,张为. 一种改进DeeplabV3网络的烟雾分割算法[J]. 西安电子科技大学学报, 2019, 46(6): 52-59.

WANG Ziyi,SU Yuting,LIU Yanyan,ZHANG Wei. Algorithm for segmentation of smoke using the improved DeeplabV3 network[J]. Journal of Xidian University, 2019, 46(6): 52-59.

图/表 9

图1

图2

图3

图4

图5

图6

表1

图7

表2

参考文献 15

[1]	MUHAMMAD K, AHMAD J, MEHMOOD I , et al. Convolutional Neural Networks Based Fire Detection in Surveillance Videos[J]. IEEE Access, 2018,6:18174-18183. doi: 10.1109/ACCESS.2018.2812835
[2]	ZHANG Q X, LIN G H, ZHANG Y M , et al. Wildland Forest Fire Smoke Detection Based on Faster R-CNN using Synthetic Smoke Images[C]//Procedia Engineering: 211. Oxford: Elsevier Ltd, 2018: 441-446.
[3]	YUAN F, ZHANG L, XIA X , et al. Deep Smoke Segmentation[J]. Neurocomputing, 2019,357:248-260. doi: 10.1109/TIP.2019.2946126 pmid: 31613768
[4]	LONG J, SHELHAMER E, DARRELL T . Fully Convolutional Networks for Semantic Segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014,39(4):640-651. doi: 10.1109/TPAMI.2016.2572683 pmid: 27244717
[5]	CHEN L C, PAPANDREOU G, KOKKINOS I , et al. Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected CRFs[C]//Conference Track Proceedings of the 2015 3rd International Conference on Learning Representations. San Diego: International Conference on Learning Representations, 2015: 149801.
[6]	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 pmid: 28463186
[7]	CHEN L C, PAPANDREOU G, SCHROFF F , et al. Rethinking Atrous Convolution for Semantic Image Segmentation[J/OL]. [2019-05-20]. https://arxiv.org/abs/1706.05587.
[8]	YU F, KOLTUN V, FUNKHOUSER T . Dilated Residual Networks[C]//Proceedings of the 2017 30th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 636-644.
[9]	DAI J, QI H, XIONG Y , et al. Deformable Convolutional Networks[C]//Proceedings of the 2017 IEEE International Conference on Computer Vision. Piscataway: IEEE, 2017: 764-773.
[10]	HU J, SHEN L, SUN G . Squeeze-and-excitation Networks[C]//Proceedings of the 2018 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. Washington: IEEE Computer Society, 2018: 7132-7141.
[11]	RONNEBERGER O, FISCHER P, BROX T . U-Net: Convolutional Networks for Biomedical Image Segmentation[C]//Lecture Notes in Computer Science: 9351. Heidelberg: Springer Verlag, 2015: 234-241.
[12]	ZHAO H, SHI J, QI X , et al. Pyramid Scene Parsing Network[C]//Proceedings of the 2017 30th IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 2881-2890.
[13]	赵敏, 张为, 王鑫 , 等. 时空背景模型下结合多种纹理特征的烟雾检测[J]. 西安交通大学学报, 2018,52(8):72-78.
	ZHAO Min, ZHANG Wei, WANG Xin , et al. Smoke Detection via Multi-texture Feature Exploration with Spatio-temporal Background Model[J]. Journal of Xi’an Jiaotong University, 2018,52(8):72-78.
[14]	BESBES O, BENAZZA-BENYAHIA A . A Novel Video-based Smoke Detection Method Based on Color Invariants[C]//Proceedings of the 2016 IEEE International Conference on Acoustics, Speech and Signal Processing. Piscataway: IEEE, 2016: 1911-1915.
[15]	BARMPOUTIS P, DIMITROPOULOS K, GRAMMALIDIS N . Smoke Detection Using Spatio-temporal Analysis, Motion Modeling and Dynamic Texture Recognition[C]//Proceedings of the 2014 European Signal Processing Conference. Poland: European Signal Processing Conference, 2014: 1078-1082.

算法	MPA/%	MIoU/%	时间/ms
UNet^[11]	93.99	83.43	86.38
PSPNet^[12]	95.76	88.60	108.86
DeeplabV3^[7](base)	97.22	89.04	89.86
Ours1	97.49	89.99	67.05
Ours2	97.70	90.92	67.12
Ours3	97.78	91.21	71.73

烟雾视频	Ours3		文献[13]		文献[14]		文献[15]
烟雾视频	R_TPR/ %	R_TNR/ %	R_TPR/ %	R_TNR/ %	R_TPR/ %	R_TNR/ %	R_TPR/ %
sBehindtheFence	98.26	94.60	97.20	96.27	94.72	100	94.44
sBtFence2	98.20	100	98.17	100	99.08	100	98.71
sMoky	98.85	100	99.68	100	86.23	100	99.78
sWasteBasket	99.41	100	97.18	98.36	99.89	92.60	99.29
sWindow	98.40	100	98.10	100	94.30	100	88.52
平均值	98.56	98.92	98.06	98.83	94.84	98.52	96.15