I-firenet: A lightweight CNN to increase generalization performance for real-time detection of forest fire in edge AI environments

Early detection of wildfire is critical to reduce the damage caused by delaying fire control within the golden time. A solution for detecting forest fires automatically is a deep-learning-based visual recognition system using autonomous drones. This paper proposes a new lightweight CNN(Convolutional Neural Network) model, called i-FireNet, that achieves increased classification accuracy and generalization performance for real-time detection of forest fire in the embedded computing environments such as drones. To better discern unresolved images of forest fires observed in the existing FireNet, this work presents an improved-FireNet(i-FireNet) CNN architecture by exploiting four novel design techniques, data augmentation, batch normalization, YCbCr pre-processing, and global average pooling. Experimental results show that the proposed i-FireNet increases the classification accuracy by 6.77% compared with the existing FireNet while maintaining its real-time performance in the edge device. Furthermore, the proposed architecture occupies 40% less memory space than the existing one. © ICROS 2020.