基于图像特征的煤矿火灾检测与识别研究

发布时间：2018-05-28 17:59

本文选题：火灾检测 + 模糊增强　；参考：《西安科技大学》2015年博士论文

【摘要】：火灾是矿井重大灾害之一,严重威胁着人类健康、自然环境和煤矿的安全生产。随着科技进步,火灾自动检测技术逐渐成为监测和火灾预警的重要手段。基于视频图像的火灾检测技术具有探测范围广、响应时间短、成本低、不受环境影响等优势,结合计算机智能技术可以提供更直观、更丰富的信息,对煤矿的安全生产具有重要意义。针对煤矿火灾检测过程中的图像预处理、特征提取、图像识别三个环节,本文主要完成了以下研究工作:针对煤矿井下照度低、光照不均匀的特点,在分析和比较传统图像增强处理算法的基础上,提出了基于模糊逻辑的图像增强预处理方法。该方法定义了新的模糊隶属度函数,通过快速最大类间方差法搜索阈值,实现了模糊增强阈值的自适应选取,减小了图像低灰度区域的像素信息损失,提高了运算速度。针对图像噪声大、对比度低的特点,提出了基于二维模糊划分最大熵的图像分割方法,采用模糊隶属度函数对目标与背景进行精细划分,既利用了图像的灰度信息以及空间邻域信息,而且兼顾了图像自身的模糊性,克服了传统的分割算法需要根据经验确定阈值的不足;提出采用粒子群优化算法对模糊熵函数进行优化,克服了运算速度慢的缺点,满足实时性需求。根据火灾发生早期的火焰图像特性,研究了火灾火焰的特征提取技术,结合火灾火焰的动静态特征、纹理特征、频闪特征,在图像分割的基础上分别对疑似区域及像素进行计算,以反映其面积蔓延、形体变化、边缘变化及闪烁特性,实现了火灾图像特征参数的提取;针对火灾图像的多特征选择问题,基于多特征融合的火灾图像特征选择技术,实现了初始特征集中特征冗余信息的去除,确定了火灾图像的分类特征参数,并将其作为火灾火焰和干扰源的识别依据。根据图像处理算法提取的早期火灾火焰的图像特征,构建了多特征融合的图像型火灾火焰检测模型。基于人工神经网络适合处理不完善和模糊信息的特点以及最小二乘支持向量机具有的小样本、非线性、训练效率高及高维模式优势分别进行火灾火焰图像识别,达到各个判据优势互补的目的,从而克服了传统的使用单一特征作为判据易引起误报的局限性,其中对最小二乘支持向量机的超参数选取在快速留一法的基础上,提出了采用共轭梯度法进行超参数优化,构建了FR-LSSVM模型。最后分别采用BP神经网络、最小二乘法支持向量机、FR-LSSVM和标准支持向量机进行分类实验,结果表明FR-LSSVM较其它算法有更好的稳定性、更快的运算速度和更高的识别率。基于提取的烟雾图像特征实现了实时烟雾检测。首先基于混合高斯分布对烟雾图像序列进行运动区域检测,提取出运动像素。根据烟雾扩散的动态特性结合烟雾疑似区域的面积增大与否将非烟雾目标进行滤除。基于所提取的动态特征、纹理特征利用支持向量机作为分类器通过多参数融合对烟雾图像和非烟雾图像进行识别分类。
[Abstract]:Fire is one of the major disasters in mine, which seriously threatens human health, natural environment and safe production of coal mine. With the progress of science and technology, fire automatic detection technology has gradually become an important means of monitoring and fire warning. The fire detection technology based on video image has a wide range of detection, short response time, low cost, unaffected by the environment and so on. Advantages, combined with computer intelligence technology can provide more intuitive, more abundant information, it is of great significance for the safety of coal mine production. In the process of coal mine fire detection, image preprocessing, feature extraction, image recognition three links, this paper mainly completed the following research work: low illumination in coal mine and uneven illumination On the basis of analyzing and comparing the traditional image enhancement processing algorithm, the image enhancement preprocessing method based on fuzzy logic is proposed. This method defines a new fuzzy membership function, searches the threshold by fast maximum inter class variance method, realizes the adaptive selection of the fuzzy enhancement threshold, and reduces the image of the low gray area of the image. In view of the large image noise and low contrast, the image segmentation method based on the maximum entropy of two-dimensional fuzzy division is proposed. The fuzzy membership function is used to divide the target and the background fine, and the image's gray information and the spatial neighborhood information are used, and the image itself is taken into account. Fuzziness, overcome the traditional segmentation algorithm need to determine the deficiency of the threshold according to experience; propose a particle swarm optimization algorithm to optimize the fuzzy entropy function, overcome the shortcomings of the slow operation speed, meet the real-time demand. According to the flame image characteristics of the early fire occurrence, the feature extraction technology of fire flame is studied, combined with fire. The dynamic and static characteristics, texture features and stroboscopic features of the disaster flame are calculated on the basis of the image segmentation, to reflect the area spread, the shape change, the edge change and the flicker, and the extraction of the feature parameters of the fire image, and the multi feature fusion based on multi feature fusion for the multi feature selection problem of the fire image. The fire image feature selection technique has realized the removal of the feature redundancy information of the initial feature concentration, determined the classification characteristic parameters of the fire image, and used it as the identification basis of the fire flame and the interference source. According to the image features of the early fire flame extracted by the image processing algorithm, the image fire fire with multi feature fusion was constructed. The flame detection model, based on the characteristics of artificial neural network, which is suitable for processing imperfect and fuzzy information, and small sample, nonlinear, high training efficiency and high dimensional pattern advantage of least squares support vector machine (LS SVM) to recognize the fire flame image respectively, to achieve the purpose of each criterion superiority and complement each other, thus overcoming the traditional use of a single special special. On the basis of the fast remaining one method, the super parameter optimization of the least squares support vector machine is proposed by the conjugate gradient method, and the FR-LSSVM model is constructed. Finally, the BP neural network, the least second multiplication support vector machine, the FR-LSSVM and the standard support vector are used respectively. The experimental results show that FR-LSSVM has better stability, faster operation speed and higher recognition rate than other algorithms. Based on the extracted smoke image features, real-time smoke detection is realized. Firstly, based on mixed Gauss distribution, the motion region of smoke image sequence is detected and moving pixels are extracted. The smoke diffusion is based on smoke diffusion. Based on the dynamic feature extracted, the texture features use the support vector machine as the classifier to classify the smoke image and the non smog image by multi parameter fusion.
【学位授予单位】：西安科技大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41;TD75

【相似文献】