~火灾作为一种频发的灾害事故，使得人们的生命和财产安全受到严重威胁。为避免火灾发生，对其进行早期检测已成为关键一环。传统火灾检测方法是通过传感器设备对环境中各项指标进行分析和判断，但是普遍存在检测距离受限、反应慢、误报率高等问题。随着视频监控技术的广泛应用，人工智能在视频图像处理领域中飞速发展，基于此的火焰检测方法开始形成新的突破点。但目前现有的火焰检测方法，还存在着检测准确度低、实时性差等问题。
针对现有火焰数据集清晰度低以及噪声高等问题，利用Gaussian滤波算法进行火焰图像降噪，采用基于Fast Guided Filter的暗通道去雾算法和自适应直方图均衡化算法提高火焰图像数据集质量；采用Random Erasing方法和Mosaic算法对火焰数据集进行增强，以提高训练后模型的泛化能力以及在火焰目标被遮挡情况下的模型检测准确度。
针对现有火焰检测模型准确率低且对小目标火焰存在漏检等问题，结合特征融合和注意力机制，提出一种改善对小尺寸火焰和遮挡火焰检测的优化YOLOv3网络结构，通过减少卷积后输出火焰特征信息的丢失来提高检测精度。采用标签平滑算法对样本标签进行处理，避免模型训练中“过分”相信样本标签；采用Focal Loss函数作为模型训练时的损失函数，改善火焰数据集样本不平衡分布问题。经实验验证，改进火焰检测模型平均精度为94.81%，较原始YOLOv3模型提升了3.75%。
针对模型计算量大、前向推理时间长导致的火焰检测算法实时性差等问题，首先基于BN层对稀疏化后火焰检测模型卷积层进行通道剪枝，然后利用全局通道剪枝阈值与L1范数结合的方法，对改进YOLOv3火焰检测模型Shortcut层进行剪枝，进一步减少模型层间输入和输出的推断时间。经实验验证，生成的轻量级火焰检测模型，模型大小缩减了95.8%，检测速度提高了66%。
在火焰检测精度和实时性两方面对YOLOv3模型进行了相关优化，最终进行火焰检测模型的现场实验验证，结果表明所提火焰检测方法均能很好的满足视频图像火焰检测精度和速度要求。

~As a frequent disaster, fire has seriously threatened people's lives and property safety. In order to avoid fire, early detection has become a key link. The traditional fire detection method is to analyze and judge the environmental indicators through the sensor equipment, but there are many problems, such as limited detection distance, slow response, high false alarm rate and so on. With the wide application of video monitoring technology and the rapid development of artificial intelligence in the field of video image processing, the flame detection method based on this has begun to form a new breakthrough. However, the existing flame detection methods still have some problems, such as low detection accuracy and poor real-time performance.
Aiming at the problems of low definition and high noise of existing flame image data sets, Gaussian filtering algorithm is used for flame image denoising, dark channel defogging algorithm based on fast guided filter and adaptive histogram equalization algorithm are used to improve the quality of flame image data sets, and random filtering algorithm is used In order to improve the generalization ability of the trained model and the accuracy of model detection when the flame target is occluded, erasing method and mosaic algorithm are used to enhance the flame data set.
Aiming at the problems of low accuracy of existing flame detection models and missing detection of small target flame, combined with feature fusion and attention mechanism, an improved yolov3 network structure is proposed to improve the detection of small flame and occluded flame. The detection accuracy is improved by reducing the loss of output flame feature information after convolution. The label smoothing algorithm is used to deal with the sample labels to avoid "over believing" the sample labels in the model training, and the focal loss function is used as the loss function in the model training to improve the unbalanced distribution of the samples in the flame data set. The experimental results show that the average accuracy of the improved model is 94.81%, which is 3.75% higher than that of the original model.
In order to solve the problem of poor real-time performance of flame detection algorithm caused by large amount of calculation and long forward reasoning time of the model, firstly, the convolution layer of the sparse flame detection model is pruned based on BN layer, and then the shortcut layer of the improved yolov3 flame detection model is pruned by combining the global channel pruning threshold and L1 norm, so as to further reduce the input and output errors between the model layers Infer the time. The experimental results show that the model size is reduced by 95.8% and the detection speed is increased by 66%.
In terms of flame detection accuracy and real-time performance, the yolov3 model is optimized. Finally, the flame detection model is verified by field experiments. The results show that the proposed flame detection method can well meet the requirements of video image flame detection accuracy and speed.

[1]阳婷. 基于视频监控的火灾探测系统的研究与实现[D].东华大学, 2016.

[2]赵义文, 郭瀚文. 线型感温火灾探测器现场定量检测技术研究[J].消防科学与技术, 2018, 37(11):1554-1556.

[3]李惠菁. 两种图像火灾探测技术的对比分析[J]. 消防科学与技术, 2017, 36(011):1557-1559.

[4]Du S Y, Liu Z G. A comparative study of different color spaces in computer-vision-based flame detection[J]. Multimedia Tools and Applications, 2016, 75(17):10291-10310.

[5]张丹丹, 章光, 陈西江, 等. 改进YCbCr和区域生长的多特征融合的火焰精准识别算法[J]. 激光与光电子学进展, 2020, 57(6):061022

[6]严云洋, 唐岩岩, 郭志波, 等. 融合色彩和轮廓特征的火焰检测[J].微电子学与计算机,2011,28(10):137-141+145.

[7]顾俊俊, 赵敏, 吴毅杰. 早期火灾火焰尖角计算算法的研究[J]. 青岛大学学报(工程技术版), 2010, 25(001):24-27.

[8]Tian Q. An Autoadaptive Edge-Detection Algorithm for Flame and Fire Image Processing[J]. IEEE Transactions on Instrumentation and Measurement, 2012, 61(5):1486-1493.

[9]刘辉, 张云生, 张印辉, 等. 基于灰度差分统计的火焰图像纹理特征提取[J]. 控制工程, 2013, 20(2):213-218.

[10]Qiu G Q, Liu S, Cao D M, et al. Flame Recognition Based on Video Image[J]. Applied Mechanics & Materials, 2014, 687-691:3604-3607.

[11]刘伯运, 赵博, 王腾. 基于连续帧图像面积变化的火灾探测方法[J]. 消防科学与技术, 2016(12):1723-1725.

[12]Wang D C, Cui X, Park E, et al. Adaptive flame detection using randomness testing and robust features[J]. Fire Safety Journal, 2013, 55:116-125.

[13]Dimitropoulos K, Barmpoutis P, Grammalidis N. Spatio-Temporal Flame Modeling and Dynamic Texture Analysis for Automatic Video-Based Fire Detection[J]. IEEE Transactions on Circuits & Systems for Video Technology, 2015, 25(2):339-351.

[14]Zhao, Qian, Sun, et al. Flame Detection Using Generic Color Model and Improved Block-Based PCA in Active Infrared Camera[J]. International journal of pattern recognition and artificial intelligence, 2018.32(6):441-449.

[15]Kong S G, Jin D, Li S, et al. Fast fire flame detection in surveillance video using logistic regression and temporal smoothing[J]. Fire Safety Journal, 2016, 79(1):37-43.

[16]戴静, 严云洋, 范勇, 等. 基于BEMD和SVM的火焰检测算法[J]. 常州大学学报(自然科学版), 2017, 029(002):P71-77.

[17]李庆辉, 李艾华, 苏延召. 结合FCM聚类与SVM的火焰检测算法[J]. 红外与激光工程, 2014(5):335-341.

[18]Ko, ByoungChul. Wildfire smoke detection using temporospatial features and random forest classifiers[J]. Optical Engineering, 2012, 51(1): 7208-7215.

[19]Borges P, Izquierdo E. A Probabilistic Approach for Vision-Based Fire Detection in Videos[J]. IEEE Transactions on Circuits & Systems for Video Technology, 2010, 20(5):721-731.

[20]Xuan Truong T, Kim J M. Fire flame detection in video sequences using multi-stage pattern recognition techniques[J]. Engineering Applications of Artificial Intelligence, 2012, 25(7):1365-1372.

[21]O. Maksymiv, T. Rak and D. Peleshko. Real-time fire detection method combining AdaBoost, LBP and convolutional neural network in video sequence[C]//2017 14th International Conference The Experience of Designing and Application of CAD Systems in Microelectronics (CADSM), Lviv, Ukraine, 2017, pp. 351-353.

[22]邓军, 姚涵文, 王伟峰, 等. 基于优化InceptionV1的视频火焰超像素检测方法[J]. 激光与光电子学进展, 2021, 58(2):0210004

[23]Zhong Z, Wang M, Shi Y, et al. A convolutional neural network-based flame detection method in video sequence[J]. Signal Image and Video Processing, 2018, 12(5):316-325.

[24]Xie Y, Zhu J, Cao Y, et al. Efficient Video Fire Detection Exploiting Motion-Flicker-Based Dynamic Features and Deep Static Features[J]. IEEE Access, 2020, 8:81904-81917.

[25]张鸿, 严云洋, 刘以安, 等. 基于定位置信度和区域全卷积网络的火焰检测方法[J]. 激光与光电子学进展, 2020, 57(20):196-205.

[26]Muhammad K, Ahmad J, Mehmood I, et al. Convolutional Neural Networks based Fire Detection in Surveillance Videos[J]. IEEE Access, 2018: 18174-18183.

[27]周飞燕, 金林鹏, 董军. 卷积神经网络研究综述[J].计算机学报, 2017,40(06):1229-1251.

[28]常亮, 邓小明, 周明全, 等. 图像理解中的卷积神经网络[J]. 自动化学报, 2016, 42(9):1300-1312

[29]王博, 郭继昌, 张艳. 基于深度网络的可学习感受野算法在图像分类中的应用[J]. 控制理论与应用, 2015(8):1114-1119.

[30]张杰. 基于视频图像的火灾识别算法研究[D]. 吉林大学, 2019.

[31]蒋昂波,王维维. ReLU激活函数优化研究[J]. 传感器与微系统, 2018, 37(2):50-58.

[32]Arnous A H, Biswas A, Ekici M, et al. Optical solitons and conservation laws of Kudryashov's equation with improved modified extended tanh–function[J]. Optik - International Journal for Light and Electron Optics, 2020: 56(07):165406.

[33]秦庆喜. 基于深度网络的SAR图像地物分类算法研究[D]. 2019.

[34]李旭冬, 叶茂, 李涛. 基于卷积神经网络的目标检测研究综述[J]. 计算机应用研究, 2017, 34(010):2881-2886.

[35]Ke H, D Chen, Li X, et al. Towards Brain Big Data Classification: Epileptic EEG Identification with a Lightweight VGGNet on Global MIC[J]. IEEE Access, 2018(6):14722-14733.

[36]Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift[C]//Proceedings of the 32nd International Conference on International Conference on Machine Learning. ICML, 2015(37):448-456.

[37]Shu X, Tang J, Qi G J, et al. Image Classification With Tailored Fine-Grained Dictionaries[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2018, 28(2):454-467.

[38]Lecun Y, Bottou L. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 2020, 86(11):2278-2324.

[39]Artzai Picon, Aitor Alvarez-Gila, D M S C, et al. Deep convolutional neural networks for mobile capture device-based crop disease classification in the wild[J]. Computers and Electronics in Agriculture, 2019, 161:280-290.

[40]Zhu X, Yuan J, Xiao Y, et al. Stroke classification for sketch segmentation by fine-tuning a developmental VGGNet16[J]. Multimedia Tools and Applications, 2020, 79(45):33891-33906.

[41]Szegedy C, LiuW, Jia Y Q, et al. Goingdeeperwith convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) June7-12, 2015, Boston, MA, USA, New York, IEEE, 2015:15523970

[42]郝旭政, 柴争义. 一种改进的深度残差网络行人检测方法[J]. 计算机应用研究, 2019, 36(05):295-298+310.

[43]K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition"[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, 770-778

[44]Girshick R, Donahue J, Darrell T, et al. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation. IEEE Computer Society, 2013.

[45]Wang K, Zhou W. Pedestrian and cyclist detection based on deep neural network fast R-CNN[J]. International Journal of Advanced Robotic Systems, 2019, 16(2):11052-11062.

[46]Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6):1137-1149.

[47]Moechammad S, Cahya R, Berkah A. Detecting body parts from natural disaster victims using You Only Look Once (YOLO)[J]. IOP Conference Series: Materials Science and Engineering, 2021, 1073(1):012062.

[48]Redmon J, Farhadi A. YOLO9000: Better, Faster, Stronger[J]. IEEE, 2017:6517-6525.

[49]H Zheng, Liu J, Liao M. Study on Local Optical Flow Method Based on YOLOv3 in Human Behavior Recognition[J]. Journal of Computer and Communications, 2021, 09(1):10-18.

[50]Du L, Li L, Wei D, et al. Saliency-Guided Single Shot Multibox Detector for Target Detection in SAR Images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(5):3366-3376.

[51]赵永强, 饶元, 董世鹏, 等. 深度学习目标检测方法综述[J]. 中国图象图形学报, 2020, 25(4):629-654.

[52]Wang Q, Wang Z, Li J, et al. Improvement of Faster-RCNN Detection Algorithms for Small Size Line Accessory Equipment[J]. Journal of Physics: Conference Series, 2020, 1453:012007-012017.

[53]陈金辉, 叶西宁. 行人检测中非极大值抑制算法的改进[J]. 华东理工大学学报(自然科学版), 2015, 41(03):371-378.

[54]Lin, Tsung-Y, Goyal, Priya, Girshick Ross, et al. Focal Loss for Dense Object Detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(2):318-327.

[55]Cetin.E，Computer vision based fire detection database[DB/OL]. http://signal.ee.bilkent.edu.tr/VisiFire/Demo/SmokeClips/2014/2015-12-20.

[56]CVPR Lab. at Keimyung University．Wildfire and smoke video Datebase [DB/OL]. http://cvp-r.kmu.ac.kr/2012/2015-12-20.

[57]Jia D, Wei D, Socher R, et al. ImageNet: A large-scale hierarchical image database[J]. Proc of IEEE Computer Vision & Pattern Recognition, 2009:248-255.

[58]State key lab of fire science．At University of Science and Technology of China[DB/OL]. http://staflF.ustc.edu.cn/-yfii/vsd.html

[59]孟琭, 李诚新. 近年目标跟踪算法短评——相关滤波与深度学习[J]. 中国图象图形学报, 2019, 024(007):1011-1016.

[60]肖少明, 何小海, 王正勇,等. 基于改进的方向梯度直方图的互补跟踪方法[J]. 计算机应用与软件, 2020, 37(05):211-215.

[61]He K, Jian S, Fellow, et al. Single Image Haze Removal Using Dark Channel Prior[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2011, 33(12):2341-2353.

[62]Yuan W, Yuan X, Xu S, et al. Dense Image-Matching via Optical Flow Field Estimation and Fast-Guided Filter Refinement[J]. Remote Sensing, 2019, 11(20):2410-2417.

[63]徐同莹, 彭定明, 王卫星. 改进的直方图均衡化算法[J]. 兵工自动化, 2006, 25(007):58-59.

[64]周涛,霍兵强,陆惠玲,任海玲.残差神经网络及其在医学图像处理中的应用研究[J].电子学报, 2020, 48(07):1436-1447.

[65]C Liu, H Xie, H Tao, Z ZHA, et al. Bidirectional Attention-Recognition Model for Fine-Grained Object Classification[J]. IEEE transactions on multimedia, 2020, 22(7):1785-1795.

[66]S Yun, D Han, S Chun, et al. "CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features"[C]. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), Seoul, Korea (South), 2019, 6022-6031.

[67]Hou J, H Zeng, Cai L, et al. Multi-label Learning with Multi-label Smoothing Regularization for Vehicle Re-Identification[J]. Neurocomputing, 2019, 345(14):15-22.

[68]李克文, 李新宇. 基于SENet改进的Faster R-CNN行人检测模型[J]. 计算机系统应用, 2020, 29(04):270-275.

[69]Tan Y M, Wu P, Zhou G, et al. Combining Residual Neural Networks and Feature Pyramid Networks to Estimate Poverty Using Multisource Remote Sensing Data[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13(99):553-565.