火灾危害严重，火灾检测是预防火灾的关键环节。随着火灾防治技术的不断发展，目前基于深度学习的火灾检测技术是研究的热点，它可以取得更好的检测效果。本文通过对火灾检测技术的研究，针对图像型火灾检测过程中所存在的泛化性较差、误报率较高等问题，提出一种基于深度学习的火灾检测方法。

本文在图像预处理阶段，针对传统中值滤波在较小窗口下滤除效果较差和较大窗口下对图像滤除后图像边缘信息丢失的问题，提出一种改进的滤波算法。该算法首先求取滤波窗口内不同方向上的差值，之后根据求取的差值确定与噪声滤除点相关性最高的区域，最后取该区域的中值进行输出完成滤波。实验表明本文方法图像滤波后的结果与原图像的结构相似性均高达99%以上，使图像有较好的滤波效果，峰值信噪比均高达46/dB以上，使图像边缘信息得以保存，使后续处理得到更好的效果。

在火灾区域分割阶段，针对基于传统方法中存在的分割准确率不高的问题，提出一种基于不同数据集迁移学习的全卷积神经网络模型的分割方法。该方法首先对SegNet网络模型进行改进使之成为一种轻便型的网络，之后通过将基于不同数据集的模型参数迁移到搭建好的轻便型全卷积神经网络中进行训练，通过训练与验证过程中的损失曲线和准确率曲线对比选择出最优模型进行测试，最后通过与传统分割方法对比确定了本文方法的优越性，本文提出的分割方法准确率可达93%。

在火灾图像检测阶段，针对小样本数据集卷积神经网络过拟合的问题，提出一种基于失活概率预测的Dropout网络正则化优化卷积神经网络的火灾检测算法。该算法使用改进的Dropout网络正则化对搭建好的卷积神经网络进行优化，该优化算法通过对神经网络不同卷积层设置不同的失活概率对网络进行训练，之后对测试集进行测试，获得检测准确率等数据，重复多次获得多组数据，最后利用之前得到的失活概率和检测准确率等数据对每层卷积层的失活概率进行预测，从而得到最佳失活概率完成对网络的优化。通过实验验证，本文提出的火灾检测算法检测准确率为百分之九十以上，并且有效提高了卷积网络的泛化性，同时降低了模型的误报率，对火灾检测的研究起到了一定的促进作用。

通过实验验证本文提出的改进滤波方法能滤除噪声的同时较好的保留图像细节，其结构相似性可达99%；本文提出的分割方法能较为完整的提取出火灾区域，其分割准确率可达93%；图像检测阶段本文提出的火灾检测方法检测准确率为90%以上，能快速准确的判定图像是否为火灾图像。

Fire is a serious hazard and fire detection is a key part of fire prevention. With the continuous development of fire prevention technology, fire detection based on deep learning is becoming a research hotspot, which can achieve better detection effect. In this paper, a deep learning-based fire detection method is proposed for the problems of poor generalization and high false alarm rate that exist in the process of image fire detection.

In the image pre-processing stage, this paper addresses the problems of poor filtering effect of traditional median filtering in smaller windows and loss of image edge information after filtering in larger windows. In this paper, we propose an improved algorithm to find the median gray value of the region with the highest correlation to the noise filtering point by finding the difference value on different directions in the filtering window. The experimental results show that the structural similarity of the proposed method is up to 99% or more, and the image has a better filtering effect. The peak signal-to-noise ratio is up to 46/dB or more, the image edge information can be preserved, and better results can be obtained for the subsequent process.

In the fire region segmentation stage, a method based on a full convolutional neural network with migration learning on different data sets is proposed for the problem of low segmentation accuracy with traditional methods. The method first improves the SegNet network model to make it a lightweight network, and then migrates the model parameters based on different data sets into the built lightweight FCN for training, and selects the optimal model for testing by comparing the loss curve and accuracy curve during training and validation. Compared with other methods, the accuracy of the segmentation method proposed can reach 93%.

In the fire detection stage, for the problem of overfitting of convolutional neural network with small sample data set, a fire detection algorithm based on the regularization of Dropout network with deactivation probability prediction is proposed to optimize the CNN. The algorithm uses the improved Dropout network regularization to optimize the constructed CNN, which is trained by setting different deactivation probabilities for different convolutional layers of the neural network, and then tests the test set to obtain the detection accuracy and false positive rate, and repeats several times to obtain multiple sets of data. Finally, the deactivation probability of each convolutional layer is predicted by the previously obtained deactivation probability and detection accuracy data to obtain the best deactivation probability for optimization. It is verified by experiments that the detection accuracy of the proposed algorithm is more than 90%, and it effectively improves the generalization of the convolutional network and reduces the false alarm rate, which plays a certain role in promoting the research of fire detection.

Experiments demonstrate that the proposed improved filtering method can supress noise while image detail is preserved, and its structural similarity can reach 99%; the proposed segmentation method can extract the fire region more completely, and its segmentation accuracy can reach 93%; the detection accuracy of the proposed fire detection algorithm is more than 90%, and it can determine whether the image is a fire image quickly and accurately.

[1] Yan B, Li J, Zhang M, et al. Raman distributed temperature sensor with optical dynamic difference compensation and visual localization technology for tunnel fire detection[J]. Sensors (Basel, Switzerland), 2019, 19(10): 2320-2323.

[2] Rowland J H, Yuan L, Thomas R A. Evaluation of carbon monoxide and smoke sensors at a low ventilation velocity[J]. Mining Metallurgy & Exploration, 2021, 38(01): 603-608.

[3] Huang Y, Tsou C. A silicon-based photosensor with photosensitive cavity structure[J]. IEEE Photonics Technology Letters, 2021, 33(04): 217-220.

[4] Li P, Yang Y, Zhao W, et al. Evaluation of image fire detection algorithms based on image complexity[J]. Fire Safety Journal, 2021, 121(03): 103306.

[5] Krizhevsky A, Sutskever I, Hinton G. Image net classification with deep convolutional neural networks[C]//Conference and Workshop on Neural Information Processing Systems. Red Hook: Curran Associates Inc, 2012: 1097-1105.

[6] He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. Los Alamitos: IEEE Computer Society, 2016: 1-12.

[7] Sun Y, D Liang, Wang X, et al. DeepID3: Face recognition with very deep neural networks[J]. Computer Science, 2015, 36(08): 1-5.

[8] Tian Y D. A simple analysis of AlphaGo[J]. Acta Automatica Sinica, 2016, 42(05): 671-675.

[9] Jie Z A, Gc A, Sh A, et al. Graph neural networks: A review of methods and applications[J]. AI Open, 2020, 1(01): 57-81.

[10] Cetin A E, Dimitropoulos K, Gouverneur B, et al. Video fire detection: review [J]. Digital Signal Processing, 2013, 23(06): 1827-1843.

[11] Phillips I, Shah M, Lobo N D V. Flame recognition in video[J]. Pattern Recognition Letters, 2000, 23(01): 319-327.

[12] Chakraborty I, Paul T K. A hybrid clustering algorithm for fire detection in video and analysis with color based thresholding method[M]. Toronto: Rheumatol, 2010: 386-393.

[13] 张增, 王兵, 伍小洁, 等. 无人机森林火灾监测中火情检测方法研究[J]. 遥感信息, 2015, 30(01): 107-110+124.

[14] 陈培豪, 肖铎, 李晨辉. 多传感器融合的火灾识别算法研究[J]. 消防科学与技术, 2020, 39(06): 810-813.

[15] Töreyin B U, Dedeoğlu Y, Gűdűkbay U, et al. Computer vision based method for real-time fire and flame detection[J]. Pattern Recognition Letters, 2006, 27(01): 49-58.

[16] Qi X, Ebert J. A computer vision based method for fire detection in color videos[J] International Journal of Imaging, 2009, 2(09): 22-34.

[17] Chen J, He Y, Wang J. Multi-feature fusion based fast video flame detection[J]. Building and Environment, 2010, 45(05): 1113-1122.

[18] 贾阳, 王慧琴, 胡燕, 等. 基于改进层次聚类和SVM的图像型火焰识别[J]. 计算机工程与应用, 2014, 50(05): 165-168+194.

[19] 熊昊, 李伟. 基于SVM的视频火焰检测算法[J]. 传感器与微系统, 2020, 39(01): 143-145+149.

[20] Maria J S, Alexandra M, Miguel Almeida. Wildfire detection using transfer learning on augmented datasets[J]. Expert Systems with Applications, 2020, 142(05): 1-15.

[21] Qin Y Y, Cao J T, Ji X F. Fire detection method based on depthwise separable convolution and YOLOv3[J]. International Journal of Automation and Computing, 2021, 18(02): 300-310.

[22] Henrique de Almeida Pereira Gabriel, Minoro Fusioka Andre, Tomoyuki Nassu Bogdan,Minetto Rodrigo. Active fire detection in Landsat-8 imagery: A large-scale dataset and a deep-learning study[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 178(01): 171-186.

[23] Byoungjun K, Joonwhoan L. A video-based fire detection using deep learning models[J]. Applied Sciences, 2019, 9(14): 1-16.

[24] Akhloufi M A, Tokime R B, Elassady H. Wildland fires detection and segmentation using deep learning[C]//Pattern Recognition and Tracking. Bellingham: International Society for Optics and Photonics, 2018: 106490B.

[25] 李杰, 邱选兵, 张恩华, 等. 李传亮.基于卷积神经网络的火灾识别算法[J].计算机应用, 2020, 40(S2): 173-177.

[26] Zhong Z, Wang M J, Shi Y K, et al. A convolutional neural network-based flame detection method in video sequence[J]. Signal, Image and Video Processing, 2018, 12(08): 1619-1627.

[27] 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.

[28] Zhang C W, Yue X Y, Wang R, et al. Study on traffic sign recognition by optimized Lenet-5 algorithm[J]. International Journal of Pattern Recognition and Artificial Intelligence, 2020, 34(01):21-36.

[29] Tao C, Zhang J, Wang P. Smoke detection based on deep convolutional neural networks[C]// 2016 International Conference on Industrial Informatics. Piscataway: IEEE, 2016: 150-153.

[30] Yuan F, Shi J, Xia X, et al. High-order local ternary patterns with locality preserving projection for smoke detection and image classification[J]. Information Sciences, 2016, 372: 225-240.

[31] 吴雪, 宋晓茹, 高嵩, 等. 基于数据增强的卷积神经网络火灾识别[J]. 科学技术与工程, 2020, 20(03): 1113-1117..

[32] Luo Y, Zhao L, Liu P, et al. Fire smoke detection algorithm based on motion characteristic and convolutional neural networks[J]. Multimedia Tools and Applications, 2017, 77(12): 1-18.

[33] Wang Z, Wang Z, Zhang H, et al. A novel fire detection approach based on CNN-SVM using tensorflow[J]. Springer, Cham, 2017, 10363: 682-693.

[34] Hu Y, Lu X. Real-time video fire smoke detection by utilizing spatial-temporal convnet features[J]. Multimedia Tools and Applications, 2018, 77(22): 29283-29301.

[35] Hubel D H, Wiesel T N. Receptive fields, binocular interaction and functional architecture in the cat's visual cortex[J]. The Journal of physiology, 1962, 160(01): 106-154.

[36] Fukushima K, Miyake S. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position[J]. Biological Cybernetics, 1980, 36(04): 193-202.

[37] LeCun Y, Boser B, Denker J S, et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Computation, 1989, 1(04): 541-551.

[38] LeCun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied to document recognition[J]. Proceedings of The IEEE, 1998, 86(11): 2278-2324.

[39] Nair V, Hinton G. E, Rectified linear units improve restricted boltzmann machines[C]. International Conference on International Conference on Machine Learning. Madison: Omnipress, 2010: 807–814.

[40] Hinton G E, Srivastava N, Krizhevsky A, et al. Improving neural networks by preventing co-adaptation of feature detectors[J]. Computer Science, 2012, 3(04): 212-223.

[41] Szegedy C, Wei L, Jia Y, et al. Going deeper with convolutions[J]. IEEE Computer Society, 2014, 1(03): 1-12.

[42] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. Computer Science, 2014, 12(01): 1-14.

[43] Zhang C W, Yue X Y, Wang R, et al. Study on traffic sign recognition by optimized Lenet-5 algorithm[J]. International Journal of Pattern Recognition and Artificial Intelligence, 2020, 34(01): 21-34.

[44] Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014, 39(04): 640-651.

[45] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional networks for biomedical Image segmentation[J]. Springer International Publishing, 2015, 40(11): 234-241.

[46] Badrinarayanan V, Handa A, Cipolla R. SegNet: A deep convolutional encoder-decoder architecture for robust semantic pixel-wise labelling[J]. Computer Science, 2015, 12(07): 1-10.

[47] Wu H, Zhang J, Huang K, et al. FastFCN: Rethinking dilated convolution in the backbone for semantic segmentation[J]. Computer Science, 2019, 19(03):118-129.

[48] Kamiyama M, Taguchi A. Color conversion formula with saturation correction from HSI color space to RGB color space[J]. IEICE Transactions on Fundamentals of Electronics Communications and Computer Sciences, 2021, 13(05):23-28.

[49] Lu C T, Lin J A, Chang C Y, et al. Recognition of film type using HSV features on deep-learning neural networks[J]. Journal of Electronic Science and Technology, 2020, 18(01): 31-41.

[50] Ding C, Tang L, Cao L, et al. Preprocessing of multi-line structured light image based on radon transformation and gray-scale transformation[J]. Multimedia Tools and Applications, 2020, 80(05): 7529-7546.

[51] 杨先凤, 蒋欣岑, 杜晶晶. 融合伽马变换及分数阶的低照度图像增强算法[J]. 计算机工程与设计, 2021, 42(03): 762-769.

[52] Li Z. Contrast limited adaptive histogram equalization[J]. Computer Knowledge and Technology, 2010, 12(05): 23-34.

[53] Nodes T, Gallagher N C J. Median filters: Some modifications and their properties[J]. IEEE Transactions Acoustics Speech Signal Processing, 1982, 30(05): 739-746.

[54] 艾加秋, 王非凡, 杨兴明, 等. 基于背景匀质性双边滤波的SAR图像斑点噪声抑制算法[J]. 遥感学报, 2021, 25(05): 1071-1084.

[55] 王煜, 谢政, 朱淳钊, 等. 基于3次B样条小波变换的改进自适应阈值边缘检测算法[J]. 计算技术与自动化, 2021, 40(01): 101-103+118.

[56] 石野, 王晓凯, 刘慧锋. 基于局部滤波模板的双边滤波算法研究[J]. 测试技术学报, 2021, 35(01): 49-53.

[57] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. Computer Science, 2014, 13(09): 1-13.

[58] Zhou B C, Han C Y, Guo T D. Convergence of stochastic gradient descent in deep neural network[J]. Acta Mathematicae Applicatae Sinica, 2021, 37(01): 126-136.

[59] 张德, 李国璋, 王怀光, 张峻宁. 位姿估计自适应学习率的改进[J]. 电子测量与仪器学报, 2019, 33(06): 51-58.

[60] Kingma D, Ba J. Adam: A method for stochastic optimization[J]. Computer Science, 2014, 22(09): 1-15.

[61] 陈琼, 徐洋洋, 陈林清. 不平衡数据的迁移学习分类算法[J]. 华南理工大学学报(自然科学版), 2018, 46(01): 122-130.

[62] 杨甜甜, 郭大波, 孙佳. 基于多残差网络的遥感图像语义分割方法[J]. 测试技术学报, 2021, 35(03): 245-252.

[63] Rodriguez-Gonzalez A, Torres-Nino J, Valencia-Garcia R, et al. Using experts feedback in clinical case resolution and arbitration as accuracy diagnosis methodology[J]. Computers in Biology &Medicine, 2013, 43(08): 975-986.

[64] 薛君蕊, 王昱潭, 曲爱丽, 等. 基于改进FCN-8s的灵武长枣图像分割方法[J]. 农业工程学报, 2021, 37(05): 191-197.

[65] Zhao D, Liu L, Yu F H, et al. Chaotic random spare ant colony optimization for multi-threshold image segmentation of 2D Kapur entropy[J]. Knowledge-Based Systems, 2020, 216(15): 1-15.

[66] 张晓琳, 张冲, 杨涛. 基于改进布谷鸟算法的火焰图像阈值分割算法[J]. 微电子学与计算机, 2017, 34(01): 66-70.

[67] Basar S, Id O R, Id Z, et al. Unsupervised color image segmentation: A case of RGB histogram based K-means clustering initialization[J]. Public Library of Science, 2020, 15(10): 1-15.

[68] Zhou Z H. Ensemble methods: Foundations and algorithms[M]. Taylor & Francis, 2012.

[69] Dae-Hyun, Lee, Sang, et al. Fire detection using color and motion models[J]. IEIE Transactions on Smart Processing & Computing, 2017, 6(04): 237-245.

[70] 梅建军, 张为. 基于ViBe与机器学习的早期火灾检测算法[J]. 光学学报, 2018, 38(07): 60-67.

[71] Shang C, Zhang D, Yang Y. A gradient-based method for multilevel thresholding[J]. Expert Systems with Applications, 2021, 175(01): 114845.

[72] 唐阳山, 徐忠帅, 黄贤丞, 等. 基于Roberts算子的车道线图像的边缘检测研究[J]. 辽宁工业大学学报, 2017, 37(06): 383-386+390.

[73] Yang Y, Wei L. Grey relevancy degree and improved eight-direction Sobel operator edge detection[J]. Journal of Signal and Information Processing, 2021, 12(2): 43-55.

[74] Xu P, He Z, Qiu T, et al. Quantum image processing algorithm using edge extraction based on Kirsch operator[J]. Optics Express, 2020, 28(9): 12508-12517.

[75] Gong L H, Tian C, Zou W P, et al. Robust and imperceptible watermarking scheme based on Canny edge detection and SVD in the contourlet domain[J]. Multimedia Tools and Applications, 2020, 80(01): 439-461.

[76] 赵丽君, 周永军, 汤小红, 等. 无人驾驶深度学习模型组合剪枝算法[J]. 传感器与微系统, 2021, 40(03): 127-129.

[77] 崔颖, 施丹丹, 徐泽龙, 等. 卷积神经网络在核小体定位识别中的应用[J]. 哈尔滨工程大学学报, 2021(05): 1-7.

[78] Eunmi R, Jewon K, Jieun L, et al. Automated detection of surface automated detection of surface with thermal-structural behaviors of fire damaged concrete beams[J]. International Journal of Concrete Structures and Materials, 2020, 14(03): 317-328.

[79] 郭心悦, 胡沁涵, 刘纯平, 等. 基于迁移学习和批归一化的菜肴图像识别方法[J]. 计算机应用与软件, 2021, 38(03): 124-133.

[80] 姜山, 丁治明, 朱美玲, 等. 面向时空图建模的图小波卷积神经网络模型[J]. 软件学报, 2021, 32(03): 726-741.

[81] 田桂丰, 单志龙, 廖祝华, 等. 基于Faster R-CNN深度学习的网络入侵检测模型[J]. 南京理工大学学报, 2021, 45(01): 56-62.

[82] 陈建平, 王春雷, 王雪冬. 基于CNN神经网络的煤层底板突水预测[J]. 中国地质灾害与防治学报, 2021, 32(01): 50-57.

[83] Robert A S, Kwaku A, Francis G, et al. Hardware module design and software implementation of multisensor fire detection and notification system using fuzzy logic and convolutional neural networks[J]. Journal of Engineering, 2020, 2020: 1-16.

[84] 安凤平. 优化非线性激活函数-全局卷积神经网络的物体识别算法[J]. 小型微型计算机系统, 2021, 42(02): 393-398.

[85] Hicintuka J P, zhou W N. Video based fire detection systems on forest and wildland using convolutional neural network[J]. Journal of Donghua University, 2019, 36(02): 149-157.

[86] 陈君夫, 付章杰, 张卫明, 等. 基于深度学习的图像隐写分析综述[J]. 软件学报, 2021, 32(02): 551-578.