图像烟雾识别的成分分离算法

发布时间：2018-10-09 18:34

【摘要】：基于视频图像烟雾检测技术相对基于传感器原理的烟雾探测技术具有受环境影响小、响应速度快以及检测结果直观等优势,是实现火灾早期预警的重要手段。火灾发生初期,通常伴随着烟雾的产生,本文的主要研究工作是通过监控场景来判断是否有烟雾产生。现有的烟雾识别方法均是从图像中直接提取烟雾的可视化特征,所提取的特征包含背景和烟雾两部分信息,致使无法有效描述烟雾的特征,从而影响了烟雾识别的精度。针对该问题,本文从成像原理的角度出发,认为一幅图像是由背景图像与烟雾图像线性混合而成的,提出烟雾线性表达模型及其最优化问题,并提出成分分离算法求解该问题。成分分离算法,即通过从当前图像中单独分离出烟雾成分后,提取其纹理特征进而实现烟雾识别。该算法以矩形块为计算单位,根据相邻像素间像素具有相似性,构建局部平滑模型;同时,从整个纹理结构的角度看,纯烟雾图像位于一个低维的子空间,并且可以采用主成分分析来确定纯烟雾图像的子空间进而描述纯烟雾图像,因此,构建主成分模型。通过采用这两个模型对图像进行成分分离,得到纯烟雾成分,然后利用LBP算子提取其纹理特征,最后代入支持向量机分类器中判断烟雾是否存在,从而实现烟雾识别。通过合成图像以及真实视频数据对本文提出的算法性能进行评估,从检测精度方面与Toreyin和Tian烟雾识别算法进行对比分析。实验结果表明,该算法在室内、室外、背景复杂的情况下均能有效识别出烟雾,正检率均在93%以上,误检率以及漏检率均在在4%以下;同时,针对全覆盖浓烟、全覆盖薄烟、局部覆盖度大于50%的烟、局部覆盖度小于50%的烟四类不同类型的烟雾图像,本文算法平均精度高于Toreyin和Tian烟雾识别算法,平均检测精度为91.3%,误检率为5.4%,漏检率为3.3%。
[Abstract]:Compared with the smoke detection technology based on sensor principle, the video image smoke detection technology has the advantages of less environmental impact, fast response and intuitive detection results. It is an important means to achieve early warning of fire. In the early stage of fire, smoke is usually accompanied by smoke. The main work of this paper is to judge whether smoke is produced by monitoring scene. The existing methods of smoke recognition are to extract the visual features of smoke directly from the image. The extracted features include background and smoke information, which can not effectively describe the characteristics of smoke, thus affecting the accuracy of smoke recognition. In this paper, from the angle of imaging principle, we think that an image is a linear mixture of background image and smoke image, propose a smoke linear representation model and its optimization problem, and propose a component separation algorithm to solve the problem. The component separation algorithm is to extract the texture feature of the smoke component from the current image and then to realize the smoke recognition by separating the smoke component separately from the current image. The algorithm takes the rectangular block as the unit of calculation and constructs a local smoothing model according to the pixel similarity between adjacent pixels. At the same time, from the point of view of the whole texture structure, the pure smog image is located in a low-dimensional subspace. And the principal component analysis can be used to determine the subspace of the pure smoke image and then describe the pure smoke image. Therefore, the principal component model is constructed. By using these two models to separate the components of the image, the pure smoke components are obtained, and then the texture features are extracted by using the LBP operator. Finally, the smoke is judged in the support vector machine classifier to realize smoke recognition. The performance of the proposed algorithm is evaluated by synthetic images and real video data, and compared with Toreyin and Tian smoke recognition algorithms in terms of detection accuracy. The experimental results show that the algorithm can effectively identify smoke in indoor, outdoor and complex background, the positive detection rate is over 93%, the false detection rate and missed detection rate are below 4%. The average accuracy of this algorithm is higher than that of Toreyin and Tian smoke recognition algorithms. The average detection accuracy is 91.3%, the false detection rate is 5.4%, and the missed detection rate is 3.3%.
【学位授予单位】：昆明理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】