摘要：Fire detection in buildings is crucial for people's lives and property. Conventional temperature and smoke sensors have many disadvantages: the limited cover range; detection delays; the difficulty in distinguishing smoke and fire. Recently, research on convolutional neural networks (CNN) for fire image detection has become a hot topic. However, existing fire classification and object detection methods are often interfered with by flashlights, red objects and the high-brightness background, resulting in a high false alarm rate. Besides, light and lamps often exist in buildings. To address this issue, this paper focuses on introducing scene prior knowledge and causal inference mechanisms to suppress the lamp disturbance. Firstly, we train the YoloV3 network to detect and recognize lamps. Secondly, to reduce the dataset bias, we mask the lamp regions with the proposed Local Grabcut segmentation method. Last, compared with direct fire classification methods, our proposed methods reduce about 34.6% false alarm rate based on InceptionV4 networks. The experimental results verify the effectiveness among different CNN architectures (Resnet101, Firenet, Densenet121). The code is online at https://github.com/kailaisun/fire-detection-without-lamp.
