二分粒度聚类和KELM的全景图像火焰识别研究

Research on Panoramic Image Flame Recognition Based on the Bisecting Granular Clustering and KELM

针对现有灭火机器人视觉系统的窄视野且检测结果受光照变化干扰的问题,提出了一种应用于大视角全景图像火焰识别且抗光照变化干扰的二分粒度聚类优化的核极限学习机方法。首先,对全景图像建立抗光照变化干扰的颜色模型;然后在该颜色模型下利用经过二分和粒度思想改进的K-means聚类算法分割疑似火焰区域与非火区域;最终提取疑似火焰区域的颜色分量等特征参数作为输入向量来训练核极限学习机(KELM)分类器以提取火焰区域。经仿真研究证明,该算法能快速准确识别全景火焰图像,对光照变化具有良好的鲁棒性,且通用性强。

Aiming at the narrow field of view of the existing fire-extinguishing robot vision system and the detection results are disturbed by the illumination variation,a flame recognition method for large-angle panoramic image and resistant to changes in illumination is proposed.It is kernel limit learning machine method for bisecting granular clustering optimization.Firstly,a color model that is resistant to illumination variation is established for the panoramic image.Then,using the K-means clustering algorithm improved by bisecting and granularity,the suspected flame region and non-fire region are segmented under the color model.Finally,the color component feature parameters of the suspected flame region are extracted as input vectors to train a kernel extreme learning machine(KELM)classifier to identify the flame region.Simulation results show that the algorithm has quickly and accurately effect on the processing of panoramic images with fire flame and good robustness with respect to the illumination variation,as well as stronger universality.