基于改进Faster R-CNN的石化装置火灾增强识别方法

Enhanced Identification Method of Petrochemical Plant Fire Based on Improved Faster R-CNN

为提高复杂石化装置场景下火灾发现速度和识别准确率,特别是强化对石化厂区监控图像内小微火焰、烟雾等事故初期关键目标信息的发现能力,以应对石化装置火灾事故演化所具有的典型多米诺效应,在事故发展初期快速准确发现异常、控制事故态势发展、降低事故危害,建立了基于深度学习的石化装置明火、烟雾识别方法,以改进的Faster R-CNN作为目标信息检测算法,利用K-Means++聚类方法优化锚框尺寸,结合基于递归策略的多尺度特征融合设计,显著改善了模型小微关键目标识别能力,实验测试数据表明,改进后的模型检测平均准确率(mAP)提升11.0%,小目标平均识别准确率(mAP(s))提升19.7%。

In order to improve the fire detection speed and recognition accuracy in the complex petrochemi-cal plant scene,and strengthen the extraction ability of key target information such as flame and smoke in the monitoring image of the petrochemical plant,a fire and smoke recognition method based on deep learning was established.The improved faster R-CNN was used as the target information detection algo-rithm.The K-Means++clustering method was used to optimize the size of the anchor frame,and com-bined with the multi-scale feature fusion design based on recursive strategy,the recognition ability of key targets of microenterprises was significantly improved.Experimental test data showed that the improved mod-el detection mAP was increased by 11.0%,and the small target mAP(s)was increased by 19.7%.