Detect Fire Using Weighted Box Fusion Algorithm

Frequent fires pose a grave risk to both lives and property. Detecting these fires accurately is crucial for effective mitigation efforts. In the past, scholars have attempted to detect them by using infrared camera to capture fire flames. However, infrared thermal images of flames often suffer from unclear flame features and low pixel resolution. Moreover, it is unlikely to expect that an algorithm can extract all image features conducive to fire flame detection. Consequently, algorithms frequently misidentify similar flame-like objects, leading to a notable reduction in recognition accuracy. To address this issue, this paper proposes an integrated model W3YAD for fire detection. The model integrates three weakly supervised learning models (i.e., YOLOX, Deformable DETR, and Autoassign) that are trained using a limited number of image samples. The fusion of the three models' prediction results is achieved using the Weighted Box Fusion (WBF) algorithm to generate new fused prediction boxes. The experiments have shown that this integrated model fully harnesses the strengths of the three individual models, resulting in higher fire detection accuracy.