Generating fire objects and few-shot learning optimization method for fire situation detection model

In South Korea, there are more than 40,000 fire accidents every year, and national preventive measures are necessary because of the large number of casualties caused by building fires. For this purpose, several attempts have been made to detect fire in advance. In order to detect fire detection through an image rather than a sensor area, large amounts of data are required to learn detection model. However, in special cases such as fire, the collection of these data is limited. In this paper, we propose a method to learn the detection model by using this limited data as self-generated or synthesized learning data. The fire situation is detected through three models: a model for detecting fire factor objects to synthesize fire, a model for synthesizing / generating a fire on factor objects, and a model for detecting fire situation based on the generated data. Based on the proposed method, the experimental results show that the detection accuracy is like that of the model with large amount of data with only small amount of data, and the learning time is shortened because it is learned with small amounts of data. Therefore, the proposed method in this research shows that it is possible to learn sufficiently accurate models with only limited learning data.