A Slight Smoke Perceptual Network

Abstract

Detecting smoke from visual sensors is crucial for fire warnings. However, visually detecting smoke is still challenging because the chrominance, shape, transparency, texture, motion, and so on, of smoke vary over a wide range. A deep neural network (DNN) could detect smoke with higher accuracy, but the models are too large to run on limited resource platforms. To make the models smaller and improve the accuracy, three novel strategies are proposed in this paper: 1) feed the networks with blocks rather than the original image; 2) feed the network with condensed data rather than videos to extract dynamic characteristics; and 3) the framework has two stages, the first stage focuses on every block to find thin smoke, and the second stage focuses on the ascending and expanding motions in videos. It is not an end-to-end network, and every subnetwork works on a small matrix, even the one at the second stage, which operates on the whole image. Therefore, the network based on these strategies, the slight smoke perceptual network (SSPN), is small enough to be integrated on a raspberry pi. The experiments demonstrate that SSPN outperforms existing traditional methods and the methods based on DNN and achieves the highest accuracy and the highest sensitivity.

Keywords

• Smoke detection
• fire alarm
• deep neural network
• convolution neural network
• recurrent neural network
• dynamic characteristic