Light-FireNet: an efficient lightweight network for fire detection in diverse environments

作者：Khudayberdiev, Otabek; Zhang, Jiashu; Abdullahi, Sani M.^*; Zhang, Sheng

来源：MULTIMEDIA TOOLS AND APPLICATIONS, 2022, 81(17): 24553-24572.

DOI：10.1007/s11042-022-12552-5

• Convolutional neural networks (CNNs)
• Fire detection
• Lightweight
• Image classification
• Network design
• Deep learning
• Diverse environments

Fire and smoke detection using deep learning have recently proven to be a robust and efficient detection approach in contrast to traditional vision-based techniques. Efforts are made by researchers to leverage this promising direction but are always faced with a trade-off between performance accuracy and model size. To tackle this, we present Light-FireNet, an enhanced lightweight, fast, and cost-effective system based on a combination of lighter convolution mechanisms inspired by Hard Swish (H-Swish), and a novel architectural design built from scratch. Experimental results and performance analysis reveal that our proposed method has 32% fewer parameters than AlexNet, 3.03 MB lighter than MobileNetV2, and achieves a better detection accuracy of 97.83%, which is higher than most existing fire detection techniques in the literature.

• 单位
  西南交通大学