This study introduces a novel ensemble neural network for the classification and detection of fire and smoke in fire incidents. The proposed ensemble neural network efficiently detects smoke and promptly localizes fire incidents by addressing three key characteristics. First, the ensemble neural network effectively monitors fire incidents with a neural network for fire and smoke classification. This neural network aims to classify fire, smoke, or normal events by analyzing entire images with a low false alarm rate (FAR). Second, the ensemble neural network detects smoke or fire by activating distinct neural networks based on classification results. Each network for smoke or fire detection employs specialized architectures tailored to the target objects, i.e., fire or smoke, ensuring high detection accuracy and robustness. Third, the real-time performance of the ensemble neural network is significantly enhanced through half-tensor and Cython transformations, achieving a remarkable 69.4% improvement in inference speed. Extensive experiments with private and public image sets for fire and smoke classification and detection demonstrate that the proposed framework achieves high accuracy and a low FAR. An in-depth comparative study confirms that the proposed ensemble neural network outperforms other state-of-the-art neural networks, reducing the FAR by over 40% and achieving a fire and smoke classification accuracy of 99.8% and a detection accuracy of 96.8% mean average precision. The proposed ensemble neural network contributes to preventing property losses and casualties in fire incidents by being deployed in a fire monitoring and suppression system.
Do et al. (Tue,) studied this question.