采用机器学习的火焰前景提取算法

发布时间：2018-03-30 06:33

本文选题：火灾检测　切入点：前景提取　出处：《西安交通大学学报》2017年08期

【摘要】：针对现有火焰前景提取算法提取前景不完整、轮廓失真严重、对复杂环境适应性差等问题,提出一种采用机器学习的火焰前景提取算法。该算法结合使用监督学习方法和无监督学习方法,训练了两级专用的分类器用于确定疑似目标区域;根据Real AdaBoost分类器的输出结果计算聚类算法的初始中心,并使用计算出的聚类中心对目标区域进行聚类分割,以得到最终的前景区域。实验结果表明:该算法对强光环境、夜间环境、静态或动态干扰环境等复杂场景均具有较好的适应性,得到的前景提取误差率在2%~28%之间,低于现有其他算法,且帧运算耗时小于50ms,能够很好地完成多种场景下的火焰前景提取工作,为图像型火灾检测系统中后续的特征提取与识别奠定了基础。
[Abstract]:Aiming at the problems of incomplete foreground extraction, serious contour distortion and poor adaptability to complex environment, a flame foreground extraction algorithm based on machine learning is proposed.The algorithm combines supervised learning method with unsupervised learning method, and trains a two-level special classifier to determine the suspected target area, and calculates the initial center of the clustering algorithm according to the output of Real AdaBoost classifier.The calculated clustering centers are used to segment the target regions to obtain the final foreground regions.The experimental results show that the algorithm has good adaptability to complex scenes such as strong light environment, night environment, static or dynamic interference environment, and the error rate of foreground extraction is between 2% and 28%, which is lower than that of other existing algorithms.The frame computation time is less than 50ms, which can well complete the flame foreground extraction under various scenarios, which lays a foundation for the subsequent feature extraction and recognition in the image-based fire detection system.
【作者单位】：天津大学电子信息工程学院;南开大学电子信息与光学工程学院;公安部天津消防研究所火灾物证鉴定中心;
【基金】：天津市科技支撑计划重点资助项目(14ZCZDSF00020)
【分类号】：TP181;TP391.41

【参考文献】