Most ships use a sensor-based detection method to detect a fire. This method uses several sensors to detect flame, smoke, and heat to determine if there is a fire or not. This method provides only the approximate location of fire in the event of fire detection. Most sensors are also installed on the ceiling of each floor. This can take from several seconds to several minutes because flame, smoke, and heat must reach the sensor to detect a fire.
In contrast, image-based fire detection can detect fires quickly and safely over long distances. In the case of image-based fire detection, various information such as the location, degree and size of the fire can be easily obtained, and it is possible to detect the fire very quickly in order to judge the presence or absence of fire through the video.
Early image-based fire detection systems consisted of several complex algorithms. First, the original video entered through the camera is pre-treated, and then it goes through a feature extraction process such as edge, line, corner, circle, and texture. It also has a process of analyzing extracted feature information and interpreting it for purpose. Here too, the characteristics of the work are different depending on whether it is a still image or a video, and the processing methods and algorithms vary depending on the purpose. The detection performance was also limited by the experimental environment and data characteristics.
In recent years, with the performance of GPU improved due to the development of computer technology, deep learning technology has begun to develop, and this deep learning technology has been applied to image-based detection. This is a much better way to perform without having to go through the traditional pre-treatment process and feature extraction process.
In this paper, we evaluate the fire detection performance by applying the YOLO algorithm, which is one of the deep learning algorithms, to image-based fire detection. The video were filmed through a simulated fire experiment, assuming the fire situation in the engine room of the training ship SAENURI and SAEYUDAL. After converting the filmed videos into images, they were divided into images for learning and tests. Using YOLO, learned smoke and flames in one class called Fire. Based on the learning results, fire detection performance was evaluated using test images. The detection rate, false detection rate, and accuracy rate were 0.9791%, 0.0153%, and 0.9819%, respectively. In the evaluation, the detection time per test image was about 7ms.
So, Experiments have shown that it was detected quickly and accurately compared to a sensor-based fire detector.