Simple and Efficient Smoke Segmentation Based on Fully Convolutional Network

Mao-shen LIU, Xiao-tian XIE, Ke GU, Jun-fei QIAO


In this paper, a shallow fully convolutional network for image smoke segmentation is designed to solve the real-time monitoring of smoke emitted by the flare stack. This algorithm can quickly and effectively distinguish the smoke area in the image, which can determine the actions of the flare stack control system to improve combustion efficiency. The main difficulty in segmenting the smoke in the flare stack image is the variegated texture and shape of the smoke and the varying brightness, color and other disturbances of the background. According to the above problems, we only use one layer of convolution to extract large numbers of low level features such as texture and color, and further utilize two special convolutional layers, separable convolution and 1×1 convolution, to map the final segmentation result. Through experiments on different data sets, our algorithm has the best accuracy and efficiency.


Fully convolutional network, Smoke segmentation, Flare stack smoke.


Full Text:



  • There are currently no refbacks.