Data Labeling Research for Deep Learning Based Fire Detection System

Abstract

Due to the various sizes and shapes of fire flame and smoke, detection using deep learning is not easy. Fire flame and smoke data collection is an important part of deep learning. Data labeling for training has a significant impact on deep learning results. This paper describes a study on the data labeling method used in fire detection systems using deep learning. There are several programs for objects labeling. However, the labeling result (coordinate of the object) is used for data training in the same way. Fire flames and smoke vary in size over distance. In addition, since the shape varies, the results vary depending on the labeling method, which is a significant difficulty in fire flame and smoke detection. In this paper, we experimented by dividing the size of label into two types. First, the collected fire flame and smoke images are labeled big and small size. Second, based on the labeled results, training is divided into three (big, small, and mixed). Experimental results showed that the use of big size was higher than the others, but the small size of fire flame or smoke was missed. Therefore, it is considered that the labeling for the fire detection must be made not too large and not too small to increase the detection rate. And if we apply the appropriate size to special purpose, the result will be good.