Degree Name

Master of Research


School of Electrical, Computer and Telecommunications Engineering


In the past, research on fire detection has paid much attention to improvement of fire detection hardware equipment, especially the upgrade of sensors. However, the detection method based on deep neural networks is rarely mentioned. Especially the neural networks that can be embedded into the limited storage space modules.

This research uses light-weighted neural networks based on MobileNet-V2 to perform classification experiments on an expanded fire detection dataset. In the experiment, the network structure was quantitatively analyzed, and the classification results were compared and analyzed. The research results are as follows:

  • The improved light-weighted convolutional neural networks based on MobileNet-V2 further reduce the model size, and the feature extraction capability of the improved model is not significantly reduced compared with the original structure.
  • The modified neural network based on Squeeze-Net is another strategy for building a lightweight backbone for fire detection. The modified backbone is combined with 1×1 convolutional layers to reduce the number of parameters and keep the same size output of the module.
  • In the improved light-weighted neural networks, the modified channel attention block and spatial attention block are combined into a lightweight mixed attention module, then embedded into the backbone to enhance the feature expression ability of the neural networks.

The research on lightweight neural networks has enriched the methods for fire detection and provided more strategies for applying neural networks to embedded devices.

FoR codes (2008)




Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.