结合注意力机制的火灾检测算法

鉴于现有的火灾检测手段大多依赖于感温探测器和感烟探测器,但感温探测器和感烟探测器的探测具有一定的滞后性,无法实时准确地检测出初期火灾的问题,因此,构建了一个大规模多场景的火灾图像数据集;同时对图像数据集进行了火焰和烟雾目标标注,并提出了一种具有注意力机制的火灾检测算法,采用颜色分析的方法检测出图像中火焰和烟雾的疑似区域;再对火焰和烟雾目标的疑似区域进行关注,通过结合深度网络的特征提取能力,得到火灾目标的检测模型;实验结果表明,此方法在检测火灾任务上取得了更优的效果,相比于基于YOLOv3的火灾检测模型,mAP(mean average precision)提高了5.9%,同时满足了实时检测的需求。

Fire detection algorithm combined with attention mechanism

In view of the fact that most of the existing fire detection methods rely on heat detectors and smoke detectors, but the detection of temperature detectors and smoke detectors has a certain hysteresis, and cannot accurately detect the initial fire problems in real time. , Constructed a large-scale multi-scene fire image dataset, and at the same time annotated the flame and smoke object of the image dataset, and proposed a fire detection algorithm with attention mechanism, which uses color analysis to detect the suspected areas of flames and smoke, and then pay attention to the suspected areas of flames and smoke object. By combining the feature extraction capabilities of the deep network, the fire object detection model is obtained; the experimental results show that this method achieve better fire detection tasks Compared with the mean average precision (mAP) of the fire detection model based on YOLOv3, the effect is improved by 5.9%, while meeting the needs of real-time detection.