【作者】 张海波；

【导师】 赵运基；

【作者基本信息】 河南理工大学 ， 控制科学与工程， 2020， 硕士

【摘要】 森林火灾的发生严重威胁人民生命财产安全,生态安全。森林火灾的及时发现及预警对保护人民生命财产安全,生态安全具有重要意义。随着图像采集设备的广泛应用,基于图像处理的火焰或烟雾检测方法被广泛用于森林火灾检测。基于图像处理的火焰或烟雾检测方法效率取决于特征提取,准确的特征描述有助于提升检测算法的精度,降低误报率和漏报率。本文针对火焰或烟雾特征的有效提取与高效检测问题开展研究,设计了三种基于图像特征优化的火焰或烟雾检测算法:(1)传统的颜色特征应用于火焰描述时,颜色分量固化导致特征表征不充分,易受相似背景干扰,鉴于此,提出了一种自适应颜色空间的森林火灾检测算法。该算法采用CN构建自适应三通道颜色空间模型,依据主元特征确定火焰候选区域,对候选区域应用基于核的支持向量机(KSVM)分类算法进行二次识别排除干扰。实验结果表明,应用多颜色空间模型求取主元从而确定火焰候选区域,然后应用KSVM对候选区域进行二次识别的方法能够精准定位火灾区域,降低误报率。(2)基于卷积神经网络(CNN)的火焰检测算法中,待检图像直接输入识别网络影响识别速度,不适合应用于在线检测场合,是因为图像帧中存在火焰的概率通常较低。传统灰度拟合确定待检区域的方式,疑似区域过多将进一步影响识别效率。鉴于此,提出了一种三通道拟合的卷积神经网络森林火灾识别算法。通过对RGB三通道图像进行拟合,以优化特征,依据三通道拟合图像构建候选火焰Mask区域,提取候选区域送入CNN识别。为了解决训练CNN过程中,非线性叠加导致目标函数优化速度慢的问题,运用PCA求取特征向量初始化卷积核。实验结果表明,应用三通道拟合提取火焰候选区域送入CNN识别的方法,实现在线精准火点定位,提升检测精度和速度。(3)传统的深度网络模型是基于模式识别方法,因此,将传统的深度网络应用于森林火灾烟雾检测中时,多通道的深度特征存在通道冗余,影响检测算法的检测效率。鉴于此,提出了一种通道感知策略以删减冗余通道,优化特征表达,以提升检测算法的速度和精度。同时将深度可分离算法引入烟雾检测模型,在确保检测精度的前提下,进一步提升检测算法的检测速度。实验结果表明,与无通道删减的烟雾检测算法相比,提出的检测算法在不降低准确率的前提下,提升检测算法的检测速度,同时深度可分离机制进一步提升了算法的检测速度。更多还原

【Abstract】 The occurrence of forest fires seriously threatens the safety of lives and the property of people,and ecological security.Early warning and real-time detection of forest fires are great importance for protecting safety of lives and property of people and ecological safety.With the widespread application of image acquisition equipment,flame or smoke detection methods of forest fires based on the image processing are widely used in forest fires detection.The high-performance smoke or flame detection techniques depend on valid feature extraction.Accurate feature descriptions help improve the precision of the detection algorithm and reduce the false positive rate and false negative rate.This paper focuses on the valid feaure extraction and detection of flame or smoke features,and designed three flame or smoke detection algorithms based on image feature optimization.(1)When traditional color features are applied to describe flame feature,the components of traditional color space are solidified,color feature is insufficiently characterized,and it is susceptible to similar background interference.In order to solve that,the adaptive color space forest fire detection algorithm was proposed.The algorithm used the CN(Color Names)multi-color space model to rebuild an adaptive three-channels color space model.Flame candidate regions were selected based on the principal component features.The candidate regions were identified according to applying Support Vector Machine based on Kernel function(KSVM)algorithm.Finally,output the recognition result.The simulation experiment results show that the multi-color space model is used to obtain principal components to confirm the candidate area,and then the KSVM is used to identify the candidate regions to accurately locate the fire area and reduce the false alarm rate.(2)In the flame detection algorithm based on two-dimensional Convolutional Neural Network(CNN),gray-scale function intergrated images resulted in feature expression insufficiency,and because the probability of flames in the image fram is usually low,directly inputting the image to the recognition network affects the detection speed,and is unsuitable for occasion of online flame detection.In view of that,a three-channel fitting convolutional neural network forest fire recognition algorithm was proposed.By integrating the RGB three-channel images,the flame mask region of the candidate target is constructed according to caculate contrast,and the candidate region is extracted,then sent to CNN for recognition.In order to solve the problem of slow the optimization speed of objective function which caused by non-linear superposition during back propagation of CNN training,the Principal Component Analysis(PCA)was applied to obtain the feature vectors to initialize the convolution kernel.The experimental results show that three-channel feature integration method effectually extract flame candidate regions and send to CNN recognition,achieve accurate online fire flame localization,and improve the iteration speed and detection speed.(3)The traditional deep learning network is based on the pattern recognition method.Therefore,when traditional deep network is applied to forest fire smoke detection,the existence of channel redundancy in the multi-channel deep features,which affects the efficiency of detection algorithm.The Target-Aware algorithm was porposed to eliminate channel redundancy and optimize feature expression to enhance speed and precision of detection algorithm.At the same time,Depthwise separable algorithm was introduced into the smoke detection model,and speed of the detection algorithm is further improved while ensuring the detection accuracy.Experiment results show that compared with the smoke detection without channel pruning,the Target-Aware improves detection speed and precision.The Depthwise Separable further improves the detection speed.更多还原