基于改进Deeplabv3+的烟雾区域分割识别算法

doi:10.12305/j.issn.1001-506X.2021.02.06

摘要/Abstract

摘要：

烟雾遮挡使基于图像的寻的制导系统目标识别困难,如何提高该类区域分割识别准确性、降低虚警率是一个亟待解决的课题。现有Deeplabv3+算法对烟雾分割时存在漏分割、错分割等问题,细节损失严重,整体分割精度低。本文提出基于改进Deeplabv3+模型的烟雾区域分割算法,创新异感受野融合的基于空洞卷积的金字塔构型(atrous spatial pyramid pooling, ASPP)结构,进一步扩大空洞卷积感受野,降低信息损失带来的不良影响;优化骨干网络,加入多尺度融合模块,降低网络参数量和计算量;引入通道注意力模块,强化对重点通道的特征学习能力,提高模型训练速度和分割精度。实验结果表明,改进Deeplabv3+模型在测试集中平均交并比为91.03%,分割效率为12.64帧/秒,分割效果远远优于传统模式识别算法;与Deeplabv3+基础模型相比,以较小的检测效率损失为代价取得了更高的分割精度,全场景理解和细节处理能力均有明显提升。

关键词: 语义分割, 深度学习, 烟雾, 识别算法

Abstract:

Smoke occlusion has an adverse effect on the target recognition of image-based homing guidance system. It is an urgent problem to improve the accuracy of smoke region segmentation recognition accuracy and reduce the false alarm rate. Some problems exist in Deeplabv3+basic algorithm such as missing and wrong segmentation, which result in serious loss of details and low overall segmentation accuracy. This paper proposes a smoke region segmentation algorithm based on improved Deeplabv3+model. An atrous spatial pyramid pooling (ASPP) structure of different-sensory field fusion based on the empty convolution is proposed to further expand the empty convolution receptive field and reduce the adverse impact of information loss. The backbone network is optimized and multi-scale fusion module is added to reduce the network parameters and calculation. The channel attention module is introduced to strengthen the feature learning ability of key channels, speed up the model training and improve the model segmentation accuracy. The experimental result shows that the improved Deeplabv3+model has an mean intersection over union (MIoU) ratio of 91.03% in the test set and a segmentation efficiency of 12.64 FPS, which has better result of segmentation than the traditional pattern recognition algorithm. The improved Deeplabv3+model achieves higher segmentation accuracy at the cost of less detection efficiency loss compared with the Deeplabv3+basic model, and the abilities of full scene understanding and detail processing are significantly improved.

Key words: semantic segmentation, deep learning, smoke, recognition algorithm

中图分类号:

刘志赢, 谢春思, 李进军, 桑雨. 基于改进Deeplabv3+的烟雾区域分割识别算法[J]. 系统工程与电子技术, 2021, 43(2): 328-335.

Zhiying LIU, Chunsi XIE, Jinjun LI, Yu SANG. Smoke region segmentation recognition algorithm based on improved Deeplabv3+[J]. Systems Engineering and Electronics, 2021, 43(2): 328-335.

图/表 15

图1

图2

图3

图4

表1

图5

图6

表2

图7

图8

图9

表3

表4

表5

表6

参考文献 25

1	HUANG L , SONG J M , ZHANG M Q , et al. Optical flow based guidance system design for semi-strapdown image homing guided missiles[J]. Chinese Journal of Aeronautics, 2016, 29 (5): 1345- 1354. doi: 10.1016/j.cja.2016.08.011
2	齐航, 袁健全, 李磊, 等. 基于深度学习的红外烟幕区域分割技术[J]. 控制与信息技术, 2019, (4): 18- 22, 57.
	QI H , YUAN J Q , LI L , et al. A method of smoke area segmentation for infrared images based on deep learning[J]. Control and Information Technology, 2019, (4): 18- 22, 57.
3	YUAN J , YUAN W , JIA Y , et al. Renovated identifying method of the active infrared smoke based on the texture feature analysis[J]. Journal of Safety and Environment, 2016, 16 (2): 86- 89.
4	SONG W , WANG Y P , ZHANG W J , et al. Real-time smoke detection in substation based on color and motion features[J]. Transducer and Microsystem Technologies, 2015, 34 (12): 127- 130.
5	YUAN F N , SHI J T , XIA X . High-order local ternary patterns with locality preserving projection for smoke detection and image classification[J]. Information Sciences, 2016, 372 (1): 225- 240.
6	YE W , ZHAO J H , WANG S , et al. Dynamic texturebased smoke detection using surfacelet transformand HMT model[J]. Fire Safety Journal, 2015, 73 (4): 91- 101.
7	NAYAK S R , MISHRA J . A modified triangle box-counting with precision in error fit[J]. Journal of Information and Optimization Sciences, 2018, 39 (1): 113- 128. doi: 10.1080/02522667.2017.1372155
8	PANIGRAHY C , AYAN S , NIHAR K M . Image texture surface analysis using an improved differential box counting based fractal dimension[J]. Powder Technology, 2020, 364 (15): 279- 299.
9	LIU A J, LU J, ZHANG G Q, et al. Concept drift detection via equal intensity k-means space partitioning[EB/OL].[2020-5-21].https://arxiv.org/abs/2004.11587.
10	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]//Proc.of the Procedia Engineering, 2018: 441-446.
11	梁杰, 李磊. 基于深度学习的红外图像遮挡干扰检测方法[J]. 兵工学报, 2019, 40 (7): 1401- 1410. doi: 10.3969/j.issn.1000-1093.2019.07.009
	LIANG J , LI L . Infrared image occlusion interference detection method based on deep learning[J]. Acta Armamentarii, 2019, 40 (7): 1401- 1410. doi: 10.3969/j.issn.1000-1093.2019.07.009
12	FRIZZI S, KAABI R. Convolutional neural network for video fire and smoke detection[C]//Proc.of the IEEE Industrial Electronics Society, 2016: 877-882.
13	YIN Z J , WAN B Y , YUAN F N , et al. A deep normalization and convolutional neural network for image smoke detection[J]. IEEE Access, 2017, 5, 18429- 18438. doi: 10.1109/ACCESS.2017.2747399
14	XU G , ZHANG Y M , ZHANG Q X , et al. Video smoke detection based on deep saliency network[J]. Fire Safety Journal, 2019, 105 (9): 277- 285.
15	YUAN F N , ZHANG L , XIA X , et al. Deep smoke segmentation[J]. Neurocomputing, 2019, 357 (3): 248- 260.
16	汪梓艺, 苏育挺. 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 Z Y , SU Y T . Algorithm for segmentation of smoke using the improved DeeplabV3 network[J]. Journal of Xidian University, 2019, 46 (6): 52- 59.
17	LIANG M, HU X L. Recurrent convolutional neural network for object recognition[C]//Proc.of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 3367-3375.
18	LI X Q , CHEN Z X , WU Q M , et al. 3D parallel fully convolutional networks for real-time video wildfire smoke detection[J]. IEEE Trans.on Circuits and Systems for Video Technology, 2018, 30 (1): 89- 103.
19	LONG J , SHELHAMER E , DARRELL T . Fully convolutional networks for semantic segmentation[J]. IEEE Trans.on Pattern Analysis & Machine Intelligence, 2014, 39 (4): 640- 651.
20	RONNEBERGER O, FISCHER P, BROX T. U-Net: convolutional networks for biomedical image segmentation[EB/OL].[2020-05-22]. https://arxiv.org/abs/1505.04597.
21	CHEN L C, PAPANDREOU G, SCHROFF F, et al. Rethinking atrous convolution for semantic image segmentation[EB/OL]. [2020-5-22].https://arxiv.org/abs/1706.05587.
22	CHEN L C, ZHU Y, PAPANDREOU G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[EB/OL].[2020-5-22].https://arxiv.org/abs/1802.02611.
23	杨波,陶青川,董沛君.改进Deeplab v3+网络的手术器械分割方法[EB/OL].[2020-05-22]. http://kns.cnki.net/kcms/detail/11.2127.TP.20200221.0832.004.html.
	YANG B, TAO Q C, DONG P J, et al. Surgical instrument segmentation method based on improved Deeplabv3+network[EB/OL]. Computer Engineering and Applications: 1-7[2020-05-22]. http://kns.cnki.net/kcms/detail/11.2127.TP.20200221.0832.004.html.
24	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proc.of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
25	HU J, SHEN L, SUN G. Squeeze-and-excitation Networks[C]// Proc.of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2018: 7132-7141.

扩张率r	感受野width	有效运算元素num	信息利用率η/%
6	13	9	5.33
12	25	9	1.44
18	37	9	0.66
12+6	37	49	3.58
18+12	61	81	2.18

属性名称	参数值
形状	球形
半径/μm	0.03~0.05
质量/μg	0.1~0.3
颜色	白色
透明度	0.8~1.5
寿命/s	0.5~1.5
初始位置/mm	按需
初始速度/(mm/s)	按需

算法	MIoU/%			每秒传输帧数	内存MB
算法	可见光	红外	总体	每秒传输帧数	内存MB
改进Deeplabv3+	91.09	90.76	91.03	12.64	147.3
Deeplabv3+基础模型	88.32	87.52	88.17	12.77	161.7
差分盒维数膨胀	7.15	1.25	6.06	0.11	0.005

组别	异感受野融合ASPP	骨干网络优化与多尺度融合	通道注意力模块	MIoU/%	每秒传输帧数
1	-	-	-	88.17	12.77
2	√	-	-	89.64	12.17
3	√	√	-	89.71	13.35
4	√	√	√	91.03	12.64

组别	骨干网络	MIoU/%	每秒传输帧数	内存/MB
1	本文改进模型	91.03	12.64	147.30
2	MobileNet-v2	85.58	31.32	8.54
3	MobileNet-v2(depth=0.5)	84.39	51.88	2.94
4	ResNet50	86.35	18.94	157.72
5	ResNet101	87.36	14.21	263.44