Semantic Fire Segmentation Model Based on Convolutional Neural Network for Outdoor Image

Han-Soo Choi, Myeongho Jeon, Kyungmin Song, Myungjoo Kang

Abstract: In this paper, we proposed a semantic fire image segmentation method using a convolutional neural network. The simple but powerful method proposed is middle skip connection achieved through the residual network, which is widely used in image-based deep learning. To enhance the middle skip connection, we constructed a pair of convolution layers, hereafter referred to as input convolution and output convolution, to be inserted in front and behind of the entire architecture. Conse- quently, the middle skip connection yields a stronger feedback effect compared to when only the short skip connection of the residual block and the long skip connec- tion are used. The validity of the proposed method has been confirmed by using the FiSmo dataset and the Corsican Fire Database based on various evaluation metrics. Comparative analysis shows that the proposed model outperforms previous fire seg- mentation deep learning models and image processing algorithms.