维度加权模式动态纹理特征的火焰检测

PDF下载HTML

Title:

Fire detection based on dynamic texture featuresunder a dimension-weighted mode

作者:

严云洋^1,2, 陈垂雄^1,2, 刘以安², 高尚兵¹

1. 淮阴工学院 计算机与软件工程学院, 江苏 淮安 223003;
2. 江南大学 物联网工程学院, 江苏 无锡 214122

Author(s):

YAN Yunyang^1,2, CHEN Chuixiong^1,2, LIU Yi’an², GAO Shangbing¹

1. Faculty of Computer & Software Engineering, Huaiyin Institute of Technology, Huaian 223003, China;
2. School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, China

关键词:

静态纹理; 动态纹理; 正交特征; 加权特征; 支持向量机; 火焰检测; 特征提取; 局部二值模式

Keywords:

static texture; dynamic texture; orthogonal feature; weighted feature; support vector machine; flame detection; feature extraction; local binary pattern

分类号:

TP391

DOI:

10.11992/tis.201607021

摘要:

对疑似火焰区域提取纹理特征时，用局部三值模式描述火焰静态纹理特征不利于区分火焰与其他纹理均匀的干扰物，用KNN算法（k-nearest neighbor algorithm）分类效率较低。针对这些问题，提出用三正交平面局部混合模式（three orthogonal planes local mixed pattern，LMP-TOP）描述火焰的静动态纹理，再输入维度加权的支持向量机进行分类识别。LMP-TOP是对第一维XY平面，采用八邻域的均匀局部二值模式（uniform local binary pattern，LBP^u2）三正交平面局部混合模式表示火焰的静态纹理特征；对第二维XT和第三维YT平面，则采用局部三值模式（local ternary patter，LTP）融入火焰在时间维度上的变化信息，这样在得到火焰的静态特征的同时也融入了其动态特征。根据3个维度单独用于识别的准确率，赋予其相应的权重，用维度加权的支持向量机进行分类识别。实验结果表明，相比Sthevanie等算法，本文所提出的方法火焰识别率和检测效率均较高。

Abstract:

In fire detection modeling, a local ternary pattern is generally used to extract the static and dynamic textures of the suspected flame. But it is difficult to distinguish the flame from other uniform texture interferences when a local ternary pattern is used to describe the static texture features. The efficiency is low when the KNN (k-Nearest Neighbor) algorithm is used for classification. Aimed at solving these problems, a novel method is proposed here, whereby an LMP-TOP (local mixed pattern-three orthogonal planes) method is used to depict the static and dynamic textures of a suspected flame area. A dimension-weighted support vector machine was used for the classification. Applying LMP-TOP, an eight neighborhood uniform local binary pattern (LBP^u2) was used to denote the static texture features of the flame on the 1st-dimension plane XY, and a local ternary pattern was used to describe the change in flame information on the 2nd-and 3rd -dimension planes, XT and YT respectively, by fusing with information in the time dimension. The static and dynamic characteristics of the flame were therefore integrated. The dimension weight was assigned according to the individual recognition accuracy. Then, a support vector machine with dimension weighting was used for classification. Experimental results show that the accuracy of flame identification and the detection efficiency are better with the proposed method than with corresponding algorithms such as Sthevanie.