Smoke Segmentation Method Based on Super Pixel Segmentation and Convolutional Neural Network

The steps of the fire disaster are from smokes to open flame. It has significant and practical meaning to use fixed camera stands to detect smoke. As the rapid development of AI in recent years the methods that using deep learning to monitor smoke pixels has owned technical foundation, But compared with the flame, the smoke has more complex pixel gray value. Segmentation results often affected by fog, water mist, clouds and other factors. On the other hand, as the deep learning is belong to supervision learning, we need to mark the smoke pixels before the model training. It is easy to mismark pixels due to man-made factors during the practical operation. So as to solve these problems mentioned, this paper regarded single frame with smoke as research object, using SegNet model of deep leaning to split the smoke pixel from the image, and then divide the pixels contained smoke into blocks by simple linear iterative clustering(SLIC).In the end, we combine the result of super pixel segmentation with SegNet model. The experimental results show that the results obtained by the above method is better than the original SegNet and the details of the smoke can be better reflected.