ScholarWorks@성균관대학교

초록

With the growth of automated surveillance and intelligent safety systems, rapid and accurate fire detection technologies have become essential. Recent advances in computer vision and deep learning have made vision-based fire detection increasingly popular due to their accuracy and non-contact nature. However, in the deep-learning-based system, the discrepancy between the input image resolution and the model input size can limit inference accuracy. To address this, we propose a technique maintaining aspect ratio while adapting image resolution to match the model input size, preserving crucial fire and smoke information. This method allows immediate use of pretrained models, facilitating rapid deployment in real-world scenarios. The simulation results show that the proposed inference method can improve prediction accuracy and enable the construction of a reliable fire detection system.

키워드