Lightweight and Efficient Deep Learning Model for Fire Detection

Forest fires seriously threaten ecosystems, the environment, and human life. Research and efforts to prevent forest fires with early detection systems have become the main focus lately. Early detection of forest fires requires real-time and accurate performance to extinguish the fire immediately. An accurate Artificial Intelligence method for detection is using deep learning. However, the drawback of deep learning is the long inference process due to the large size of the model. The dataset that we use in this research consists of two classes: the fire class and the non-fire class. In this research, we propose using quantization to shrink deep learning models. We use quantization to convert floating point 32-bit to 16-bit. The deep learning model that we use is the CNN DenseNet-121 model. The quantization results on the deep learning model show that the values for macro precision, macro recall, macro F1, and accuracy have not changed from the original deep learning, namely 0.99, 0.99, 0.99, and 0.99. The size of the deep learning model has decreased from 28.2 MB to 13.7 MB. It affects the inference model time, which is 36 times faster using quantization of deep learning than the original deep learning model. The speed of inference model quantization reaches 0.062 s per image. Through these results, the quantization result in the deep learning model can run in real-time and accurately.