YOLO-ULNet: Ultralightweight Network for Real-Time Detection of Forest Fire on Embedded Sensing Devices
Lei Huang, Zhuoyue Ding, Cheng Zhang, Run Ye, Bin Yan, Xiaojia Zhou, Wenbo Xu, Jinhong Guo
Abstract
Forest fires can cause tremendous damage to the environment and irreparable loss to individual properties. Using deep learning algorithms on embedded sensing devices to detect smoke and flame for real-time detection of forest fire is essential to minimize the risk to the environment and individual properties. This article proposes an ultralightweight network, YOLO-ULNet, for real-time detection of forest fire on embedded sensing devices. Initially, a lightweight network, YOLO-LNet, consisting of input, backbone, neck, and output head was designed to reduce the parameter quantity and computation quantity. The validation results of YOLO-LNet on the computer showed that the [email protected], speed, computation quantity, and parameter quantity were 71.2%, 110 FPS, 0.7 GFLOPs, and 0.26 M, respectively, representing a performance that was superior to that of YOLOv5s. The test results of YOLO-LNet on the Raspberry Pi (RPi) 4B showed that the precision and speed were 77.75% and 14.39 FPS, respectively, where the speed required for real-time detection (20 FPS) still could not be achieved. Subsequently, channel pruning and feature distillation model compression methods were employed to obtain the ultralightweight network YOLO-ULNet with improved speed and precision. The validation results of YOLO-ULNet on the computer showed that the [email protected], speed, computation quantity, and parameter quantity were 69.1%, 115 FPS, 0.4 GFLOPs, and 0.19 M, respectively. Meanwhile, the test results of YOLO-ULNet on the RPi 4B showed that the precision and speed were 74.50% and 24.57 FPS, respectively, meeting the requirements of the real-time detection of forest fire.