Citation

Hu X, Li M, Zhang D. Int. J. Veh. Des. 2023; 91(1/2/3): 87-106.

Copyright

(Copyright © 2023, Inderscience Publishers)

DOI

10.1504/IJVD.2023.131044

PMID

unavailable

Abstract

Accurately and efficiently detecting and localising the flame is critical for preventing fire disasters. However, most high-performance deep models require high hardware resources, making them hard to deploy on edge or mobile intelligent devices. This paper proposes a lightweight deep convolutional neural network, called FlameNet, for flame detection and localisation. The proposed FlameNet is derived from YOLOv4 with the following modifications: First, MobileNetV2 replaces the CSPDarknet53 as the new backbone network of YOLOv4; Second, the Coordinate Attention module is added to the inverted residual linear bottleneck of MobileNetV2; Third, the Depthwise Separable Convolutions is used in PANet of YOLOv4. There is no large-scale public dataset available, and we produced a real-world flame (RWF) dataset containing 13,129 images. Qualitative and quantitative analysis and comparisons of experimental results demonstrate that the accuracy and efficiency of the proposed FlameNet for flame detection out-perform the original YOLOv4 and the other relative models.

Language: en