Abstract: Due to the lack of flame segmentation data set, the application of traditional image segmentation methods on flame segmentation study is inadequate, and the model comparison test is not enough. To deal with these problems, based on the construction of the flame segmentation data set, 4 kinds of semantic segmentation models and 2 kinds of backbone networks which perform well in public dataset were chosen for training and testing, and were compared and analyzed under different application scenario. Experimental results showed that, U-Net model has better effect in the flame segmentation, in which U-Net+Resnet50 has the best comprehensive effect, while U-Net+Mobilenet V2 has slightly worse effect, but higher running speed.

Key words: fire protection; image processing; deep learning; neural networks; flame segmentation