E-ISSN : 2508-7894
Various techniques are being researched to effectively detect forest fires. Among them, techniques using object detection models can monitor forest fires over wide areas 24 hours a day. However, detecting forest fires early with traditional object detection models is a very challenging task. While they show decent accuracy for thick smoke and large fires, they show low accuracy for faint smoke and small fires, and frequently generate false positives for lights that are like fires. In this paper, to solve these problems, we focus on leveraging local characteristics such as contours and textures of fire and smoke, which are crucial for accurate detection. Based on this approach, we propose EDAM (Edge driven Attention Module) that performs enhancement by richly utilizing contour and texture information of fire and smoke. EDAM extracts important edge information to generate feature maps with emphasized contour and texture information, and based on this map, performs Attention Mechanism to emphasize key characteristics of smoke and fire. Through this mechanism, the overall model performance was improved, with AP_sincreasing from 0.154 to 0.204 and AP_0.5 from 0.779 to 0.784, resulting in a significant improvement in AP_Svalue to 32.47%. In practice, the model applying this technique showed excellent inference speed while greatly improving detection performance for small objects compared to existing models and reduced false positive rates for building and street light illumination in nighttime environments that are easily mistaken for fire.