火灾对人类生命和财产安全造成极大威胁。传统的温感、烟感、光感等火灾探测器检测范围有限，易受外界干扰，难以适应复杂环境要求，如森林、农田、工厂或其他室内外场所等。烟雾作为火灾初期显著的视觉特征，利用监控视频进行烟雾检测对于火灾预警具有重要意义。因此，本文以复杂场景的烟雾视频为研究对象，提出了基于传统特征分类和深度学习两种烟雾检测方法，主要工作内容如下：

（1）为了解决不同环境下烟雾视频照度低、对比度差且含有噪声等问题，提出了一种全局自适应Retinex的图像增强算法。首先，利用HSV颜色模型对S分量进行非线性拉伸处理；然后，使用融合双边滤波Retinex算法对V分量进行增强；最后，针对部分图像对比度提升不足问题，引入全局自适应对数增强算法作修正处理。实验表明，本方法可以自适应增强烟雾图像对比度，去除噪声同时较好保护了图像细节信息。对于高亮度图像，自适应函数从对数曲线近似为线性曲线，可以避免出现过增强现象。

（2）针对传统特征分类方法易受背景干扰、识别准确率偏低等问题，提出了一种多特征融合与改进随机森林的视频烟雾检测算法。首先，使用ViBe运动检测和HSV颜色模型方法确定疑似烟雾区域；然后，基于疑似区域提取烟雾的小波高频能量、烟雾增长率和动态纹理等特征，融合归一化后组成特征向量；最后，利用排序和聚类方法挑选分类能力强和相关性小的决策树，通过构建改进随机森林分类器完成烟雾有无的判断。实验表明，改进随机森林的视频烟雾检测算法在复杂环境下具有较好的识别效果，烟雾检测率达到95.2%，相比其他分类器平均提高了3%。

（3）鉴于上述传统检测方法需人工选取特征且无法获取烟雾位置信息，提出了一种基于深度学习的ECA-bneck-YOLOv4视频烟雾检测算法，在主干网络中引入ECA-bneck模块来增强模型检测性能。首先，模块将标准卷积替换成深度可分离卷积，利用逐通道卷积和逐点卷积降低网络参数量，同时减少主干网络中CSP模块的数量，以提高烟雾目标检测速率；然后，针对烟雾在复杂场景下的动态特性，使用ECA注意力机制抑制冗余背景干扰，并加入1×1卷积构成倒残差结构，增强网络特征学习能力，最终提升目标检测精度。实验表明，本方法较传统特征分类方法检测率提高了3.4%，在复杂场景下泛化能力更强，还能实现火灾定位功能。相比同类检测算法，检测帧率平均提高了9.25帧/秒，具有更高的检测速率。

Fire a great threat to human life and property safety. Traditional fire detectors such as temperature, smoke and light sensors have a limited detection range, are vulnerable to external interference and are difficult to adapt to complex environmental requirements, such as forests, farmland, factories or other indoor and outdoor places. Smoke is a prominent visual feature in the early stage of fire, and video smoke detection is of great significance for fire alarm. Therefore, this paper takes smoke videos of complex scenes as the research object, and two smoke detection methods based on traditional feature classification and deep learning are proposed. The main work contents are as follows:

(1) In order to solve the problems of low illumination, poor contrast and noise in smoke video in different environments, a global adaptive Retinex image enhancement algorithm is proposed. Firstly, the S component is nonlinear stretched by HSV color model. Then, the V component is enhanced by Retinex algorithm fused with bilateral filtering. Finally, aiming at the problem of insufficient contrast enhancement for some images, a global adaptive logarithmic enhancement algorithm is introduced for correction. Experimental results show that this method can enhance the contrast of smoke image adaptively and protect the image details while noise is suppressed. For high brightness images, the adaptive function approximates from logarithmic curve to linear curve, which can avoid over enhancement.

(2) Aiming at the traditional feature classification methods are prone to background interference and low recognition accuracy, a video smoke detection algorithm based on multi-feature fusion and modified random forest is proposed. Firstly, ViBe motion detection and HSV color model are employed to determine the complete candidate smoke area; Then, the wavelet high-frequency energy, smoke growth rate and dynamic texture of smoke are extracted based on the candidate area, and the feature vector is formed after fusion and normalization; Finally, sorting and clustering methods are used to select decision trees with strong classification ability and low correlation, and the modified random forest classifier is constructed. Experimental results show that the modified random forest video smoke detection algorithm has a good recognition effect in complex environment, and the smoke detection rate is 95.2%, which is 3% higher than that of other classifiers.

(3) In view of the above traditional detection methods need to manually select features and cannot obtain smoke location information,, an ECA-bneck-YOLOv4 video smoke detection algorithm is proposed, and the detection performance is enhanced by modifying the backbone network and introducing ECA-bneck module. Firstly, standard convolution is replaced by depth-separable convolution, and the number of network parameters is reduced by depthwise convolution and pointwise convolution. At the same time, the number of CSP modules in the backbone network is reduced to improve the detection rate. Then, arming at the dynamic characteristics of smoke in complex scenes, the ECA attention mechanism is employed to suppress the interference from redundant background, and 1×1 convolution is added to form an inverted residual structure, which enhances the feature learning ability of the network and improves the accuracy of target detection ultimately. Experimental results show that the proposed method is 3.4% higher than traditional feature classification method, the generalization ability is stronger in complex scenes, and it can also realize fire localization. Compared with similar detection algorithms, the detection frame rate is increased by 9.25 frames per second, which has a higher detection rate.

Fire a great threat to human life and property safety. Traditional fire detectors such as temperature, smoke and light sensors have a limited detection range, are vulnerable to external interference and are difficult to adapt to complex environmental requirements, such as forests, farmland, factories or other indoor and outdoor places. Smoke is a prominent visual feature in the early stage of fire, and video smoke detection is of great significance for fire alarm. Therefore, this paper takes smoke videos of complex scenes as the research object, and two smoke detection methods based on traditional feature classification and deep learning are proposed. The main work contents are as follows:

(1) In order to solve the problems of low illumination, poor contrast and noise in smoke video in different environments, a global adaptive Retinex image enhancement algorithm is proposed. Firstly, the S component is nonlinear stretched by HSV color model. Then, the V component is enhanced by Retinex algorithm fused with bilateral filtering. Finally, aiming at the problem of insufficient contrast enhancement for some images, a global adaptive logarithmic enhancement algorithm is introduced for correction. Experimental results show that this method can enhance the contrast of smoke image adaptively and protect the image details while noise is suppressed. For high brightness images, the adaptive function approximates from logarithmic curve to linear curve, which can avoid over enhancement.

(2) Aiming at the traditional feature classification methods are prone to background interference and low recognition accuracy, a video smoke detection algorithm based on multi-feature fusion and modified random forest is proposed. Firstly, ViBe motion detection and HSV color model are employed to determine the complete candidate smoke area; Then, the wavelet high-frequency energy, smoke growth rate and dynamic texture of smoke are extracted based on the candidate area, and the feature vector is formed after fusion and normalization; Finally, sorting and clustering methods are used to select decision trees with strong classification ability and low correlation, and the modified random forest classifier is constructed. Experimental results show that the modified random forest video smoke detection algorithm has a good recognition effect in complex environment, and the smoke detection rate is 95.2%, which is 3% higher than that of other classifiers.

(3) In view of the above traditional detection methods need to manually select features and cannot obtain smoke location information,, an ECA-bneck-YOLOv4 video smoke detection algorithm is proposed, and the detection performance is enhanced by modifying the backbone network and introducing ECA-bneck module. Firstly, standard convolution is replaced by depth-separable convolution, and the number of network parameters is reduced by depthwise convolution and pointwise convolution. At the same time, the number of CSP modules in the backbone network is reduced to improve the detection rate. Then, arming at the dynamic characteristics of smoke in complex scenes, the ECA attention mechanism is employed to suppress the interference from redundant background, and 1×1 convolution is added to form an inverted residual structure, which enhances the feature learning ability of the network and improves the accuracy of target detection ultimately. Experimental results show that the proposed method is 3.4% higher than traditional feature classification method, the generalization ability is stronger in complex scenes, and it can also realize fire localization. Compared with similar detection algorithms, the detection frame rate is increased by 9.25 frames per second, which has a higher detection rate.

[1] Muhammad K, Khan S, Elhoseny M, et al. Efficient Fire Detection for Uncertain Surveillance Environment[J]. IEEE Transactions on Industrial Informatics, 2019, 15(5): 3113-3122.

[2] Li J, Yan B, Zhang M, et al. Long-Range Raman Distributed Fiber Temperature Sensor With Early Warning Model for Fire Detection and Prevention[J]. IEEE Sensors, 2019, 19(10): 3711-3717.

[3] 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(3): 303-322.

[4] 李洪昌, 安明伟. 基于总有界变分的森林火灾烟雾图像检测方法[J]. 电子测量与仪器学报, 2020, 34(11): 211-217.

[5] Gaur A, Singh A, Kumar A, et al. Fire Sensing Technologies: A Review[J]. IEEE Sensors, 2019, 19(9): 3191-3202.

[6] Sun L, Cao Y, Wu W, et al. A multi-target tracking algorithm based on Gaussian mixture model[J]. Journal of Systems Engineering and Electronics, 2020, 31(3): 482-487.

[7] Li L, Wang Z, Hu Q, et al. Adaptive Nonconvex Sparsity Based Background Subtraction for Intelligent Video Surveillance[J]. IEEE Transactions on Industrial Informatics, 2020, 17(6): 4168-4178.

[8] Zhang H, Qian Y, Wang Y, et al. A ViBe based moving targets edge detection algorithm and its parallel implementation[J]. International Journal of Parallel Programming, 2020, 48(5): 890-908.

[9] Kim H, Ryu D, Park J. Smoke detection using GMM and adaboost[J]. International Journal of Computer and Communication Engineering, 2014, 3(2): 123-126.

[10] Yan H, Wang H, Qian Z, et al. Early Fire Smoke Image Segmentation in a Complex Large Space[J]. Open Construction & Building Technology Journal, 2015, 9(1): 27-31.

[11] Ye S, Bai Z, Chen H, et al. An effective algorithm to detect both smoke and flame using color and wavelet analysis[J]. Pattern Recognition & Image Analysis, 2017, 27(1): 131-138.

[12] 汪鑫, 吴开志, 俞子荣, 等. 基于PoolNet显著性和SURF-VIBE模型的林火视频烟雾提取算法[J]. 南昌航空大学学报(自然科学版), 2020, 34(2): 94-100.

[13] Deng X, Yu Z, Wang L, et al. Smoke Image Segmentation Based on Color Model[J]. Journal on Innovation and Sustainability RISUS, 2015, 6(2): 130-138.

[14] 李笋, 石永生, 汪渤, 等. 基于颜色增强变换和MSER检测的烟雾检测算法[J]. 北京理工大学学报, 2016, 36(10): 1072-1078.

[15] Prema C E, Vinsley S S, Suresh S. Multi feature analysis of smoke in YUV color space for early forest fire detection[J]. Fire Technology, 2016, 52(5): 1319-1342．

[16] Yu C, Jun R, Jinjun R, et al. Video Fire Smoke Detection Using Motion and Color Features[J]. Fire Technology, 2010, 46(3): 651-663.

[17] 刘恺, 刘湘, 常丽萍, 等. 基于YUV颜色空间和多特征融合的视频烟雾检测[J]. 传感技术学报, 2019, 32(2): 237-243.

[18] 邓实强, 丁浩, 杨孟, 等. 基于视频图像的公路隧道火灾烟雾检测[J]. 隧道建设(中英文), 2022, 42(2): 291-302.

[19] Toreyin B U, Dedeoglu Y, Cetin A E. Contour based smoke detection in video using wavelets[C]// 2006 14th European Signal Processing Conference, 2006: 1-5.

[20] Zhang Y, Wang H, Fan X. Algorithm for detection of fire smoke in a video based on wavelet energy slope fitting[J]. Journal of Information Processing Systems, 2020, 16(3): 557-571.

[21] Islam M R, Amiruzzaman M, Nasim S, et al. Smoke Object Segmentation and the Dynamic Growth Feature Model for Video-Based Smoke Detection Systems[J]. Symmetry, 2020, 12(7): 1075.

[22] Wu X, Lu X, Leung H. A Video Based Fire Smoke Detection Using Robust AdaBoost[J]. Sensors. 2018, 18(11): 3780.

[23] 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): 67-73.

[24] 丁怀对, 刘申友, 许玉坤, 等. 基于运动块追踪的视频烟雾探测方法[J]. 安全与环境学报, 2016, 16(4): 96-100.

[25] Wang G, Zhang G, Choi K, et al. Deep Additive Least Squares Support Vector Machines for Classification With Model Transfer[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2019, 49(7): 1527-1540.

[26] Liu J, Huang J, Sun R, et al. Data Fusion for Multi-Source Sensors Using GA-PSO-BP Neural Network[J]. IEEE Transactions on Intelligent Transportation Systems, 2021, 22(10): 6583-6598,

[27] Yan J, Bai X, Zhang W, et al. No-reference image quality assessment based on AdaBoost_BP neural network in wavelet domain[J]. Journal of Systems Engineering and Electronics, 2019, 30(2): 223-237.

[28] Yuan F, Fang Z, Wu S, et al. Real-time image smoke detection using staircase searching based dual threshold AdaBoost and dynamic analysis[J]. IET Image Processing, 2015, 9(10): 849-856.

[29] Dimitropoulos K, Barmpoutis P, Grammalidis N. Higher order linear dynamical systems for smoke detection in video surveillance applications[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2017, 27(5): 1143-1154.

[30] 文泽波, 康宇, 曹洋, 等. 基于随机森林特征选择的视频烟雾检测[J]. 中国科学技术大学学报, 2017, 47(8): 653-664.

[31] 房世超. 基于视频的烟雾检测[D]. 北京: 中国地质大学, 2016.

[32] Pei Y, Huang Y, Zou Q, et al. Effects of Image Degradation and Degradation Removal to CNN-Based Image Classification[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(4): 1239-1253.

[33] Zhang Y, Gao X, He L, et al. Objective Video Quality Assessment Combining Transfer Learning With CNN[J]. IEEE Transactions on Neural Networks and Learning Systems, 2020, 31(8): 2716-2730.

[34] Ye T, Zhao Z, Zhang J, et al. Low-altitude small-sized object detection using lightweight feature-enhanced convolutional neural network[J]. Journal of Systems Engineering and Electronics, 2021, 32(4): 841-853.

[35] Cao Y, Zhang W, Bai X, et al. Detection of excavated areas in high-resolution remote sensing imagery using combined hierarchical spatial pyramid pooling and VGGNet[J]. Remote Sensing Letters, 2021, 12(12): 1269-1280.

[36] Sreedhar P, Satya S, Nandhagopal N. Classification Similarity Network Model for Image Fusion Using Resnet50 and GoogLeNet[J]. Intelligent Automation and Soft Computing, 2022, 31(3): 1331-1344.

[37] Yang Q, Jiang S, Chen J, et al. Crack detection based on ResNet with spatial attention[J]. Computers and Concrete, 2020, 26(5): 411-420.

[38] Li J, Liang X, Shen S M, et al. Scale-aware Fast R-CNN for pedestrian detection[J]. IEEE Transactions on Multimedia, 2017, 20(4): 985-996.

[39] Hwang Y J, Lee J G, Moon U C, et al. SSD-TSEFFM: New SSD using trident feature and squeeze and extraction feature fusion[J]. Sensors, 2020, 20(13): 3630.

[40] Chai E, Ta L, Ma Z, et al. ERF-YOLO: A YOLO algorithm compatible with fewer parameters and higher accuracy[J]. Image and Vision Computing, 2021, 116: 104317.

[41] Li X, Chen Z, Wu Q, et al. 3D Parallel Fully Convolutional Networks for Real-time Video Wildfire Smoke Detection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2020, 30(1): 89-103.

[42] Gu K, Xia Z, Qiao J, et al. Deep Dual-Channel Neural Network for Image-Based Smoke Detection[J]. IEEE Transactions on Multimedia, 2019, 22(2): 311-321.

[43] Liu H, Lei F, Tong C, et al. Visual Smoke Detection Based on Ensemble Deep CNNs[J]. Displays, 2021, 69: 102020.

[44] Li C, Yang B, Ding H, et al. Real-time video-based smoke detection with high accuracy and efficiency[J]. Fire Safety Journal, 2020, 117: 103184.

[45] 刘通, 程江华, 华宏虎. 结合YdUaVa颜色模型和改进MobileNetV3的视频烟雾检测方法[J]. 国防科技大学学报, 2021, 43(5): 80-85.

[46] 王洋, 程江华, 刘通, 等. 一种多网络模型融合的烟雾检测方法[J]. 计算机工程与科学, 2019, 41(10): 1771-1776.

[47] 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(6): 992-996.

[48] Huo Y, Zhang Q, Jia Y, et al. A Deep Separable Convolutional Neural Network for Multiscale Image-Based Smoke Detection[J]. Fire Technology, 2022: 1-24.

[49] 刘丽娟, 陈松楠. 一种基于改进SSD的烟雾实时检测模型[J]. 信阳师范学院学报(自然科学版), 2020, 33(2): 305-311.

[50] 杜立召, 徐岩, 张为. 一种双网融合的分阶段烟雾检测算法[J]. 西安电子科技大学学报, 2020, 47(4): 141-148.

[51] 江泽涛, 覃露露, 秦嘉奇, 等. 一种基于MDARNet的低照度图像增强方法[J]. 软件学报, 2021, 32(12): 3977-3991.

[52] Drago F, Myszkowski K, Annen T. Adaptive Logarithmic Mapping For Displaying High Contrast Scenes [C]// Computer Graphics Forum, 2003, 22(3): 419-426.

[53] Kallel F, Hamida A B. A New Adaptive Gamma Correction Based Algorithm Using DWT-SVD for Non-Contrast CT Image Enhancement[J]. IEEE Transactions on NanoBioscience, 2017, 16(8): 666-675.

[54] Ojala T, Pietikainen M, Maenpaa T. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns[J]. IEEE Transactions on pattern analysis and machine intelligence, 2002, 24(7): 971-987.

[55] 高文, 汤洋, 朱明. 复杂背景下目标检测的级联分类器算法研究[J]. 物理学报, 2014, 63(09): 156-164.

[56] Perner P. How to compare and interpret two learnt Decision Trees from the same Domain[C]// 2013 27th International Conference on Advanced Information Networking and Applications Workshops, IEEE, 2013: 318-322.

[57] Yang B, Xu X, Ren J, et al. SAM-Net: Semantic probabilistic and attention mechanisms of dynamic objects for self-supervised depth and camera pose estimation in visual odometry applications[J]. Pattern Recognition Letters, 2022, 153: 126-135.

[58] Cheng D, Meng G, Cheng G, et al. SeNet: Structured edge network for sea–land segmentation[J]. IEEE Geoscience and Remote Sensing Letters, 2016, 14(2): 247-251.

[59] Chen L, Tian X, Chai G, et al. A new CBAM-P-Net model for few-shot forest species classification using airborne hyperspectral images[J]. Remote Sensing, 2021, 13(7): 1269.