火灾是一种极具破坏性的灾害，给人们的生命、财产和环境都带来了严重的威胁。火灾发生的初期会产生大量的烟雾，而在火灾中期才会有火焰的产生。所以为了提前预知火灾并尽早的采取救援措施，烟雾的提前检测显得尤为重要。针对烟雾识别过程中误报率较高以及在烟雾分割中大目标烟雾边缘和小目标分割不理想的问题，本文提出了改进的基于深度学习的烟雾识别与分割方法，旨在提高火灾预警的准确性和效率。

本文主要研究内容如下：

（1）针对当前烟雾识别算法存在误报率较高的问题，本文提出了一种结合Inception和Transformer结构的双分支烟雾识别方法TCF-Net。将卷积神经网络学习局部信息的能力与Transformer中的自注意力机制学习全局上下文信息的能力相结合，其次，通过Inception结构使卷积核种类多样化，丰富了特征种类，又减少了通道数的冗余，同时在特征提取过程中嵌入了特征耦合模块(FCU)，连续地对双分支中的局部特征和全局信息进行交互，以最大程度保留双分支中的局部信息和全局信息，提高该算法的性能。改进后的网络可以更好的提取烟雾的特征，降低了烟雾识别的误报率，并且将准确率提升至97.8%，证实了该算法有较好的性能。

（2）针对当前大多数烟雾分割算法对大目标烟雾边缘和小目标烟雾分割不理想导致精度较低的问题，本文提出了一种基于改进的DeepLabV3+的轻量化烟雾分割方法。本文将主干网络替换成了MobileNetV2，同时对空洞卷积金字塔池化模块(ASPP)进行了改进，将ASPP的空洞率设置为4、8、12、16以提高对多尺度信息的提取能力，进一步在空洞卷积模块中引入了串联结构来更好的融合多尺度特征，并且在编码部分嵌入了CBAM通道与空间注意力机制，提高了对特征融合的尺度和对小目标的关注程度。改进后的模型相较于原算法，平均交并比（mIoU）和平均像素精确度（mPA）分别提高了4.81％和2.03％。实验结果表明，与DeepLabV3+模型相比，本文方法提升了烟雾分割的速率和精确度。

（3）烟雾识别与分割系统的实现。本研究基于深度学习技术，针对烟雾分类与分割任务，设计并开发了一个系统。该系统结合了本文的两种模型，实现了对烟雾的快速、准确的分类与分割的可视化。

Fire is a very destructive disaster that poses a serious threat to people's lives, property and the environment. Fires produce large amounts of smoke in the early stages of a fire, while flames are not produced until the middle stages of a fire. Therefore, in order to anticipate fires and take rescue measures as early as possible, the early detection of smoke is particularly important. Aiming at the problems of high false alarm rate during smoke recognition and unsatisfactory segmentation of large target smoke edges and small targets in smoke segmentation, this paper proposes an improved deep-learning based smoke recognition and segmentation method, which aims to improve the accuracy and efficiency of fire warning.

The main research of this paper is as follows:

(1) Aiming at the problem of high false alarm rate of current smoke recognition algorithms, this paper proposes a two-branch smoke recognition method TCF-Net that combines Inception and Transformer structures. combines the ability of convolutional neural network to learn local information with the ability of the self-attention mechanism in Transformer to learn global contextual information, and, secondly, through the Inception structure to diversify the types of convolutional kernels, which enriches the feature variety and reduces the redundancy of the number of channels, and at the same time, the feature coupling module (FCU) is embedded in the process of feature extraction, which continuously interacts with the local features and global information in the two-branch to maximize the retention of local and global information in the two-branch to improve the performance of this algorithm. The improved network can better extract the features of smoke, reduce the false alarm rate of smoke recognition, and increase the accuracy to 97.8%, which confirms that the algorithm has better performance.

(2) Aiming at the problem that most current smoke segmentation algorithms are not ideal for large target smoke edges and small target smoke segmentation resulting in low accuracy, this paper proposes a lightweight smoke segmentation method based on the improved DeepLabV3+. In this paper, the backbone network is replaced with MobileNetV2, while the null convolution pyramid pooling module (ASPP) is improved by setting the null rate of the ASPP to 4, 8, 12, and 16 in order to improve the ability of extracting multiscale information, further introducing a tandem structure in the null convolution module for better fusion of the multiscale features and embedding the CBAM in the coding part of the channel with spatial attention mechanism in the coding part to improve the scale of feature fusion and the attention to small targets. The improved model improves the mean intersection and merger ratio (mIoU) and mean pixel accuracy (mPA) by 4.81% and 2.03%, respectively, compared to the original algorithm. The experimental results show that the method in this paper improves the rate and accuracy of smoke segmentation compared to the DeepLabV3+ model.

(3) Implementation of a smoke recognition and segmentation system. In this study, a system is designed and developed for smoke classification and segmentation tasks based on deep learning techniques. The system combines the two models in this paper to achieve fast and accurate classification and segmentation visualization of smoke.

[1]程晓舫,王瑞芳,张维农等.火灾探测的原理和方法(下)[J].中国安全科学学报,1999(02):4-8.DOI:10.16265/j.cnki.issn1003-3033.1999.02.001.

[2]史劲亭,袁非牛,夏雪.视频烟雾检测研究进展[J].中国图象图形学报,2018,23(03):303-322.

[3]张洁,吴爱国,赵萌.基于纹理特征和轮廓光流矢量的烟雾识别[J].传感器与微系统,2016,35(06):17-20.DOI:10.13873/J.1000-9787(2016)06-0017-04.

[4]Lee C Y, Lin C T, Hong C T, et al. Smoke detection using spatial and temporal analyses[J]. International Journal of Innovative Computing, Information and Control, 2012, 8(6): 1-11.

[5]冯路佳,王慧琴,王可等.基于目标区域的卷积神经网络火灾烟雾识别[J].激光与光电子学进展,2020,57(16):83-91.

[6]Yang J, Chen F, Zhang W (2008) Visual-based smoke detection using support vector machine. In: 2008 4th International conference on natural computation, Jinan, pp301–305.

[7]Zhang L, Wu J, Yuan F, et al. Smoke-Aware Global-Interactive Non-local Network for Smoke Semantic Segmentation[J]. IEEE Transactions on Image Processing, 2024.

[8]Kim S Y, Muminov A. Forest fire smoke detection based on deep learning approaches and unmanned aerial vehicle images[J]. Sensors, 2023, 23(12): 5702.

[9]Huang J, Zhou J, Yang H, et al. A small-target forest fire smoke detection model based on deformable transformer for end-to-end object detection[J]. Forests, 2023, 14(1): 162.

[10]Zheng Y, Zhang G, Tan S, et al. A forest fire smoke detection model combining convolutional neural network and vision transformer[J]. Frontiers in Forests and Global Change, 2023, 6: 1136969.

[11]Yin Z, Wan B, Yuan F, et al. A deep normalization and convolutional neural network for image smoke detection[J]. Ieee Access, 2017, 5: 18429-18438.

[12]Ghali R, Akhloufi M A. BoucaNet: a CNN-transformer for smoke recognition on remote sensing satellite images[J]. Fire, 2023, 6(12): 455.

[13]Tao H, Duan Q, Lu M, et al. Learning discriminative feature representation with pixel-level supervision for forest smoke recognition[J]. Pattern Recognition, 2023, 143: 109761.

[14]Tao H. A label-relevance multi-direction interaction network with enhanced deformable convolution for forest smoke recognition[J]. Expert Systems with Applications, 2024, 236: 121383.

[15]王文朋,毛文涛,何建樑等.基于深度迁移学习的烟雾识别方法[J].计算机应用,2017,37(11):3176-3181+3193.

[16]陈俊周,汪子杰,陈洪瀚等.基于级联卷积神经网络的视频动态烟雾检测[J].电子科技大学学报,2016,45(06):992-996.

[17]Muhammad K, Ahmad J, Mehmood I, et al. Convolutional neural networks based fire detection in surveillance videos[J]. Ieee Access, 2018, 6: 18174-18183.

[18]Zhang Q et al (2018) Wildland forest fire smoke detection based on faster R-CNN using synthetic smoke images. In: 2017 8th International conference on fire science and fire protection engineering, pp 441–446.

[19]袁梅, 全太锋, 黄俊, 等. 基于卷积神经网络的烟雾检测[J]. Journal of Chongqing University of Posts & Telecommunications (Natural Science Edition), 2020, 32(4).

[20]朱家辉, 赵志瑛, 贾静静. 基于 YUV 色彩空间的烟雾区域提取方法[J]. 电子技术与软件工程, 2021 (7): 116-117.

[21]Ma Z, Cao Y, Song L, et al. A New Smoke Segmentation Method Based on Improved Adaptive Density Peak Clustering[J]. Applied Sciences, 2023, 13(3): 1281.

[22]Xing D, Zhongming Y, Lin W, et al. Smoke image segmentation based on color model[J]. Journal on Innovation and Sustainability RISUS, 2015, 6(2): 130-138.

[23]王浩远,梁煜,张为.融合多分辨率表征的实时烟雾分割算法[J].浙江大学学报(工学版),2021,55(12):2334-2341.

[24]赵楠,王晓薇.基于运动和亮度显著性的森林烟雾分割方法[J].软件工程,2021,24(05):10-12.DOI:10.19644/j.cnki.issn2096-1472.2021.05.003.

[25]Jing T, Meng Q H, Hou H R. SmokeSeger: a transformer-CNN coupled model for urban scene smoke segmentation[J]. IEEE Transactions on Industrial Informatics, 2023.

[26]Zheng Y, Wang H, Gan X, et al. U2-Net_S Fire Smoke Segmentation Based on Multiscale Edge Fusion Convolutional Encoder and Edge Supervision[J]. 2023.

[27]Yuan F, Shi Y, Zhang L, et al. A cross-scale mixed attention network for smoke segmentation[J]. Digital Signal Processing, 2023, 134: 103924.

[28]Zhang L, Yuan F, Xia X. Edge-reinforced attention network for smoke semantic segmentation[J]. Multimedia Tools and Applications, 2023, 82(20): 31259-31284.

[29]Yuan F, Zhang L, Xia X, et al. A wave-shaped deep neural network for smoke density estimation[J]. IEEE transactions on image processing, 2019, 29: 2301-2313.

[30]Li X, Chen Z, Wu Q M J, et al. 3D parallel fully convolutional networks for real-time video wildfire smoke detection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2018, 30(1): 89-103.

[31]Yuan F, Li K, Wang C, et al. A lightweight network for smoke semantic segmentation[J]. Pattern Recognition, 2023, 137: 109289.

[32]Ma Z, Cao Y, Song L, et al. A new smoke segmentation method based on improved adaptive density peak clustering[J]. Applied Sciences, 2023, 13(3): 1281.

[33]汪梓艺,苏育挺,刘艳艳等.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(06):52-59.DOI:10.19665/j.issn1001-2400.2019.06.008.

[34]刘志赢,谢春思,李进军等.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(02):328-335.

[35]Wang Z, Yang P, Liang H, et al. Semantic segmentation and analysis on sensitive parameters of forest fire smoke using smoke-unet and landsat-8 imagery[J]. Remote Sensing, 2021, 14(1): 45.

[36]Zheng Y, Wang Z, Xu B, et al. Multi-Scale Semantic Segmentation for Fire Smoke Image Based on Global Information and U-Net[J]. Electronics, 2022, 11(17): 2718.

[37]Fukushima K. A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position[J]. Biol, Cybern, 1980, 36: 193/202.

[38]陈天鹏, 胡建文. 基于深度学习的遥感图像旋转目标检测研究综述[J]. Application Research of Computers/Jisuanji Yingyong Yanjiu, 2024, 41(2).

[39]Oloyede MO, Hancke GP, Myburgh H C. 人脸识别系统综述：最新方法和挑战[J]. 多媒体工具与应用, 2020, 79(37): 27891-27922.

[40]赖鸣姝. 基于 Transformer 的自然语言处理模型综述[J]. Artificial Intelligence and Robotics Research, 2023, 12: 219.

[41]Vaswani A, Shazeer N, Parmar N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30.

[42]Dosovitskiy A, Beyer L, Kolesnikov A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv preprint arXiv:2010.11929, 2020.

[43]Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.

[44]Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 1-9.

[45]Ioffe S, Szegedy C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[C]//International conference on machine learning. pmlr, 2015: 448-456.

[46]Szegedy C, Vanhoucke V, Ioffe S, et al. Rethinking the inception architecture for computer vision[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 2818-2826.

[47]He K, Zhang X, Ren S, et al. Deep residual learning for image recognition[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016: 770-778.

[48]Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2015: 3431-3440.

[49]Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Springer International Publishing, 2015: 234-241.

[50]Chen L C, Papandreou G, Kokkinos I, et al. Semantic image segmentation with deep convolutional nets and fully connected crfs[J]. arXiv preprint arXiv:1412.7062, 2014.

[51]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, 2017, 40(4): 834-848.

[52]Chen L C, Papandreou G, Schroff F, et al. Rethinking atrous convolution for semantic image segmentation[J]. arXiv preprint arXiv:1706.05587, 2017.

[53]Chen L C, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 801-818.

[54]Szegedy C, Ioffe S, Vanhoucke V, et al. Inception-v4, inception-resnet and the impact of residual connections on learning[C]//Proceedings of the AAAI conference on artificial intelligence. 2017, 31(1).

[55]Peng Z, Huang W, Gu S, et al. Conformer: Local features coupling global representations for visual recognition[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 367-376.

[56]Hosseini A, Hashemzadeh M, Farajzadeh N. UFS-Net: A unified flame and smoke detection method for early detection of fire in video sur-veillance applications using CNNs[J]. Journal of Computational Science, 2022, 61: 101638.

[57]Ayala A, Fernandes B, Cruz F, et al. KutralNet: A portable deep learning model for fire recognition[C]//2020 International Joint Conference on Neural Networks (IJCNN). IEEE, 2020: 1-8.

[58]Dewangan A, Pande Y, Braun H W, et al. FIgLib & SmokeyNet: Dataset and deep learning model for real-time wildland fire smoke detection[J]. Remote Sensing, 2022, 14(4): 1007.

[59]Sandler M, Howard A, Zhu M, et al. Mobilenetv2: Inverted residuals and linear bottlenecks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 4510-4520.

[60]Woo S, Park J, Lee J Y, et al. Cbam: Convolutional block attention module[C]//Proceedings of the European conference on computer vision (ECCV). 2018: 3-19.

[61]Yuan F, Zhang L, Xia X, et al. Deep smoke segmentation[J]. Neurocomputing, 2019, 357: 248-260.

[62]Zhao H, Shi J, Qi X, et al. Pyramid scene parsing network[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2881-2890