Fast smoke and flame detection based on lightweight deep neural network
Authors:
- 1 August 2020
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 1, 232-235
Abstract
The occurrence of fire disaster will cause great danger to human living environment, accurate and timely early warning can reduce or even eliminate the destructive losses caused by the smoke and flame produced by fire disaster. In recent years, thanks to the remarkable achievements of deep learning in the field of image, we propose a smoke and flame detection method based on deep learning. However, the huge amount of parameters and calculations brought by the deep neural network structure results in a large model and a slow detection speed. In this regard, we use lightweight design and channel pruning method to achieve the outstanding compression and acceleration effect, and realize the accurate and fast detection tasks of smoke and flame. Extensive experiments on the smoke and flame dataset demonstrate that after using the lightweight structure modification on the Refinedet model, it can only lose 0.4% of the mAP value while saving 87.8% parameters and 97% FLOPs. After pruning, 47.5% parameters and 43.0% FLOPs can be reduced again, and the mAP value is increased by 0.7% instead. Finally, our research results can be applied to many security detections such as smart home, power transmission and transformation lines and forest protection.Keywords
This publication has 10 references indexed in Scilit:
- Soft Filter Pruning for Accelerating Deep Convolutional Neural Networks2018
- MobileNetV2: Inverted Residuals and Linear Bottlenecks2018
- Single-Shot Refinement Neural Network for Object Detection2018
- Densely Connected Convolutional Networks2017
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksIEEE Transactions on Pattern Analysis and Machine Intelligence, 2016
- Deep Residual Learning for Image Recognition2016
- You Only Look Once: Unified, Real-Time Object Detection2016
- Going deeper with convolutions2015
- A Flame Detection Synthesis AlgorithmFire Technology, 2013
- A Fast Iterative Shrinkage-Thresholding Algorithm for Linear Inverse ProblemsSIAM Journal on Imaging Sciences, 2009