Decoupled Deep Neural Network for Smoke Detection

EasyChair Preprint no. 4942

32 pagesDate: January 30, 2021

Abstract

Smoke detection is a practical technology to protect people's lives and property. Traditional methods to detect smoke are usually based on human-crafted features, such as color, texture and shape. Although these methods do work in some cases, they are not always effective because color, texture and shape of smoke are diverse. In recent years, deep learning have achieved the most accuracy with more complex structures and more numerous parameters. However, the existing methods based on human-craft features are not accurate enough, and the ones based on deep learning often take too much computing resources. To improve the detection accuracy and reduce the computational cost, inspired by the aforementioned works, we propose a decoupled sub-network to extract color and texture separately just following the procedure of the traditional human-crafted methods. The color sub-network, consisted of several $1\times1$ convolution layers, tries to find the most suitable color model by nonlinear functions. The next sub-network, based on a series of depth-wise separable convolution layers, extracts texture features and assembles them into shape features. After integrating these features, the proposed network can comprehensively determine whether there is smoke or fire. Experimental results demonstrate that our network is compact, efficient and effective, and the decoupling trick offers a critical capability needed to catalyze widespread implementation.

Keyphrases: decoupled deep neural network, deep learning, smoke detection