Abstract: Smoke detection plays an important role in early warning of fire, so one dynamic texture recognition algorithm is proposed in this paper. Firstly,the surfacelet transform is performed on image sequences. Then a generalized Gaussian model is built for the coefficients from Surfacelet transform. The obtained model parameters are regarded as feature vector, and finally the Kullback-Leibler ( KL ) distance is used as the similarity measurement method. In experiments,three kinds of Surfacelet based smoke detectionmethods,including the use of mean and variance as feature and SVM classifier for classification;the use of mean and variance as feature and Euclidean distance as the similarity measurement method;the use of generalized Gaussian model parameters as feature and Euclidean distance as the similarity measurement tool,are implemented and used for comparison. Experimental result shows that,compared with other smoke detection methods,the new algorithm has excellent performance and lower false detection rate.

Key words: Surfacelet transform, dynamic texture, generalized Gaussian model, Kullback-Leibler ( KL ) distance, Support Vector Machine(SVM), Euclidean distance

摘要： 鉴于烟雾检测对火灾预警的重要作用,提出一种基于Surfacelet 变换的动态纹理烟雾检测算法。先对图像序列进行Surfacelet 变换,再对变换后的系数进行广义高斯建模,获得与系数相对应的模型参数作为特征,最后使用KL 距离做相似性度量。与其他3 种基于Surfacelet 变换的烟雾检测方法进行对比,包括:使用均值和方差作为特征,支持向量机进行分类;使用均值和方差作为特征,欧式距离进行相似性度量;使用广义高斯模型参数作为特征,欧式距离进行相似性度量。实验结果表明,该算法可以提高烟雾检测准确性,降低误检率,有效去除类烟运动物体的干扰。

关键词: Surfacelet 变换, 动态纹理, 广义高斯模型, KL 距离, 支持向量机, 欧氏距离

CLC Number: 

• TP391. 4