融合图像分离和特征分析的烟雾检测算法研究

发布时间：2018-07-10 14:30

本文选题：视频烟雾检测 + 前景分割　；参考：《太原理工大学》2017年硕士论文

【摘要】：火灾是一种突发性高、破坏力强的自然灾害,生产生活中用火不慎也会对人的生命财产安全造成严重威胁。火灾初期烟雾先于火焰出现,烟雾探测能够为人员疏散和扑救火灾争取更多宝贵的时间。因此,火灾烟雾探测已成为当今社会迫切需要解决的课题。目前,视频烟雾检测因其高效性、非接触性、可嵌入性等优点而成为新的研究方向。为了提升烟雾视频检测系统的高效性和可靠性,论文围绕烟雾视频检测技术的三个阶段——前景目标检测阶段、烟雾特征提取阶段、烟雾识别阶段进行研究,提出了融合图像分离和特征分析的烟雾检测算法。(1)前景目标检测:针对前景目标检测,提出了基于ICA和GBVS的烟雾图像分离检测算法,改善了传统高斯混合模型所需样本量大且采样时间跨度长这一问题。该算法在烟雾分离阶段先利用ICA烟雾前景初步分离烟雾模型得到初步烟雾前景,然后通过GBVS提取图像多通道、多尺度的底层特征得到烟雾前景显著区域,最后根据颜色和纹理的特征组合进行直方图匹配来识别烟雾。实验结果表明,该算法在烟雾前景分离阶段将ICA和GBVS相结合,有效缩减了烟雾前景区域的范围,提取的烟雾区域小而集中,ROC曲线显示该算法整体性能表现优秀。(2)烟雾特征提取:针对烟雾特征提取,提出了基于多维纹理分析的烟雾特征提取检测算法。传统烟雾视频的线性动态系统(LDS)仅采用亮度值作为图像信息,忽视了其他信息如彩色视频的RGB信息,并且由于密集采样导致计算量较大。该算法首先经过烟雾颜色过滤和背景差分预处理得到烟雾候选区域,避免出现密集采样,降低了计算量;然后在多维图像块中新增RGB和HOG特征,增加了图像块的维度;最后基于对多维图像数据的高阶分解,分析烟雾视频的动态纹理特征。由于采用滑动时间窗,可以确定画面中烟雾的确切位置和烟雾发生的具体时间。实验以检测率为评价指标,采用多元比较的方法,结果表明,该算法提高了动态纹理特征分析的可靠性,同时计算量较小。(3)烟雾识别:针对烟雾识别,提出了基于广义熵模糊神经网络的烟雾视频图像聚类检测算法。该算法首先利用基于ICA和GBVS的烟雾图像分离算法得到烟雾前景,然后提取基于高阶线性动态系统(h-LDS)的多维动态纹理特征,最后利用一种广义熵模糊神经网络对特征进行训练和分类。实验结果表明,基于广义熵模糊神经网络的烟雾视频图像聚类算法有较高的聚类正确率,且训练误差较小。
[Abstract]:Fire is a kind of natural disaster with high burst and strong destructive power. The careless use of fire in production and life will also pose a serious threat to the safety of human life and property. Smoke appears before flame in the early stage of fire, and smoke detection can buy more valuable time for evacuation and fire fighting. Therefore, fire smoke detection has become an urgent problem to be solved in today's society. At present, video smoke detection has become a new research direction because of its high efficiency, non-contact, embeddability and other advantages. In order to improve the efficiency and reliability of smoke video detection system, this paper focuses on three stages of smoke video detection technology: foreground target detection stage, smoke feature extraction stage, smoke recognition stage. A smoke detection algorithm based on ICA and GBVS is proposed. (1) foreground target detection: aiming at foreground target detection, a smoke image separation detection algorithm based on ICA and GBVS is proposed. The problem of large sample size and long sampling time span for traditional Gao Si hybrid model is improved. At the stage of smoke separation, the ICA smoke foreground model is used to obtain the initial smoke foreground, and then GBVS is used to extract the multi-channel image, and the multi-scale bottom feature is used to obtain the significant region of smoke foreground. Finally, the smoke is identified by histogram matching according to the combination of color and texture. The experimental results show that the ICA and GBVS are combined in the stage of smoke foreground separation, and the range of smoke foreground region is reduced effectively. The extracted smoke region is small and concentrated ROC curve shows that the whole performance of the algorithm is excellent. (2) smoke feature extraction: for smoke feature extraction, a multi-dimensional texture analysis based smoke feature extraction detection algorithm is proposed. The traditional linear dynamic system of smoke video (LDS) only uses luminance value as image information, neglecting other information such as RGB information of color video, and because of dense sampling, the computation is large. Firstly, smoke candidate regions are obtained by smoke color filtering and background differential preprocessing to avoid dense sampling and reduce computational complexity, then RGB and HOG features are added to multi-dimensional image blocks, and the dimension of image blocks is increased. Finally, based on the higher order decomposition of multidimensional image data, the dynamic texture features of smoke video are analyzed. Because of the sliding time window, the exact location and time of smoke in the screen can be determined. The experiment takes the detection rate as the evaluation index and adopts the method of multivariate comparison. The results show that the algorithm improves the reliability of the dynamic texture feature analysis, and the computation is small. (3) smoke recognition: for smoke recognition, A clustering detection algorithm for smoke video images based on generalized entropy fuzzy neural network is proposed. Firstly, the smoke image separation algorithm based on ICA and GBVS is used to obtain the foreground of smoke, and then the multi-dimensional dynamic texture feature based on high-order linear dynamic system (h-LDS) is extracted. Finally, a generalized entropy fuzzy neural network is used to train and classify the features. The experimental results show that the clustering algorithm based on generalized entropy fuzzy neural network has higher clustering accuracy and less training error.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】