Non-Temporal Lightweight Fire Detection Network for Intelligent Surveillance Systems

Hunjun Yang,Taeyong Kim,Bowon Lee,Hyeok Jang

doi:10.1109/access.2019.2953558

Abstract

Convolutional neural networks (CNNs) have been recently applied to tackle a variety of computer vision problems. However, because of its high computational cost, careful considerations are required to design cost-effective CNNs. In this paper, we propose a CNN inspired by MobileNet for fire detection in surveillance systems. In the proposed network, color features emphasized by the channel multiplier are extracted through depthwise separable convolution, and squeeze and excitation modules further increase the representation of the channel-wise convolution. Custom Swish is used as an activation function to limit exceedingly high weights from the effects of the channel multiplier. Our proposed network achieves 95.44% accuracy for fire detection, which is higher than those achieved other existing networks. Furthermore, the number of parameters used is 38.50% fewer than that of MobileNetV2, the smallest among other networks. We believe that using the proposed CNN, CNN-based surveillance systems could be implemented in lightweight devices without using expensive dedicated processors.

Highlights

Fires can occur anywhere, at any time, and if they are not detected early, they can cause severe damages to property and people
In order to use Convolutional neural networks (CNNs) more effectively for image analysis, we propose a network for fire detection using a channel multiplier and squeeze and excitation (SE) depthwise modules, along with a modified version of the Swish activation function
We proposed the MobileNet-Fire network, which has a lower computational cost and achieves better performance compared with existing CNNs

Summary

INTRODUCTION

At any time, and if they are not detected early, they can cause severe damages to property and people. A video-based fire detection system can inform an operator by analyzing videos from CCTVs without using heat, smoke, or flame sensors. Traditional vision-based fire detection methods use handcrafted features, such as color, motion, and texture. Habiboğlu et al [8] used texture information as a feature They divided video data into spatio-temporal blocks and extracted covariance-based features. CNN has been proven to work significantly better than traditional methods without requiring a feature extraction stage because the features are learned through the network. In order to use CNNs more effectively for image analysis, we propose a network for fire detection using a channel multiplier and squeeze and excitation (SE) depthwise modules, along with a modified version of the Swish activation function. Our experimental results demonstrated that the proposed method achieves higher accuracy than the recently proposed MobileNetV2 [26], despite having fewer network parameters and a lower computational cost

PROPOSED ALGORITHM

SE-DEPTHWISE MODULE

DATASET

ROBUSTNESS ANALYSIS

Findings

CONCLUSION