ScholarWorks@성균관대학교

초록

Fires are unpredictable disasters that can cause significant loss of life and property within a short period of time. With the advancement of deep learning, object detection-based intelligent fire detection systems have been widely developed. However, the primary objective of fire detection is not to localize fire or smoke objects in an image but to quickly determine whether a fire has occurred. In this study, we build upon a previously proposed image classification-based fire detection system and further improve its performance by simplifying the class structure. Specifically, the original four-class system (Fire, Smoke, Both, None) is restructured into a three-class system (Fire, Smoke, None) by integrating the Both class into the Fire class. This restructuring reduces model complexity and clarifies classification boundaries. The proposed three-class system was evaluated using the Twins-SVT-S model and further verified with additional backbones including Twins-SVT-B, ConvNeXtV2-Nano, and ConvNeXtV2-Tiny. Experimental results show that the new class structure leads to an average increase of 1.7 percentage points in classification accuracy across all models, without changing the inference structure. These findings demonstrate that reconstructing class composition can enhance both the performance and practicality of fire detection systems, and suggest a valuable direction for the design of efficient and reliable intelligent fire detection systems.

키워드