A Study on the Dynamic Image-Based Dark Channel Prior and Smoke Detection Using Deep Learning

The detection of smoke in a fire is a very important research topic because a large amount of carbon monoxide, which is potentially lethal, can be generated and released in the early stages of a fire. In particular, if a fire occurs in the form of smoldering combustion, it produces a glowing combustion without flames on the surface of the heat source, and the temperature is over 1000 °C. In this study, the dark channel prior, an algorithm previously used for haze removal, is used to detect areas where smoke may exist. The dark channel characteristic makes it possible to effectively detect the smoke area included background interference or noise. Additionally, in order to detect the characteristic that the smoke generated from the fire rises due to the density difference at high temperatures, the area of the smoke was detected using the optical flow technique based on the Lucas–Kanade method. Image pre-processing using the dark channel prior and the optical flow technique can effectively detect the smoke areas and significantly reduce false positive rate. Through this, in order to accurately determine the filtered region as smoke or non-smoke, a Convolutional Neural Network was employed. As a result, it was confirmed that accuracy and precision were improved by 4% and 7%, respectively, compared to object detection models that performed detection without image pre-processing.