A Deep Learning Method based on SRN-YOLO for Forest Fire Detection
Li Y, Zhixi Shen, Junbei Li, Zanlin Xu
Abstract
Sensor-based methods and vision-based methods with manual features selection have been widely used in forest fire detection. However, they have limited detection areas and are only effective in some specific scenarios. In this paper, SRN-YOLO, an improved version of YOLOv3, is proposed to reach a higher accuracy and lightweight network structure for forest fire detection. First, a sparse residual network (SRN) is proposed, which solves the potential overfitting problem by modifying the skip-connection to the sparse-connection. Then, based on the sparse-connection, the residual units in the classic YOLOv3 are replaced by the sparse residual modules, and the backbone network Darknet-53 is improved to obtain the proposed SRN-YOLO network. Through comparison experiments with other YOLO-based networks such as YOLO-LITE and Tinier-YOLO, the results show that the proposed network in this paper is effective and lightweight, and can achieve higher accuracy for forest fire detection.