engleski

Modern CNNs Comparison for Fire Detection in RGB Images

Every year, fire causes thousands of deaths as well as billions of dollars of material damage. Prevention and early fire detection have become a topic of interest for many scientists. While there are many existing solutions such as smoke detectors, flame detectors, chemical sensors, infrared thermal cameras and many other hybrid systems, computer vision techniques that use raw RGB image as an input have emerged as fast, reliable, precise, and economical enough to be widely used with a satisfactory accuracy. For that purpose, Convolutional Neural Networks (CNNs) were considered as they can take input image from an RGB camera, learn its features and classify it as fire or non-fire. Another important thing to consider is their ability to be used on hardware with a limited amount of computational power, e.g. embedded systems. In this paper, four different versions of MobileNet, four versions of ResNet, and four versions of EfficientNet were evaluated by comparing their ability to detect fire while also taking into consideration their need for computational power. The evaluation was performed on a custom dataset that contains over 60, 000 images. Overall, ResNet showed the lowest performance which was somewhat expected as it is the oldest network. MobileNets and EfficientNets showed similar performance proving themselves to be capable when used as a fire detection classifiers. Also, due to their low number of parameters and low computational need, they are suitable for use in systems with limited resources.

convolutional neural network (CNN) ; deep learning ; fire detection ; image classification ; performance evaluation

nije evidentirano