Squeezed fire binary segmentation model using convolutional neural network for outdoor images on embedded device

Kyungmin Song, Han-Soo Choi, Myungjoo Kang

Abstract: Even though image-based prediction of fire events is widely used, the current predictive methods are difficult to implement due to low performance and high specifications. In this work designed to overcome such problems, we propose binary semantic segmentation for fire images by employing deep learning that can be applied to embedded devices such as Jetson TX2. To reduce the parameters and consequently the model size while maintaining the performance, we replaced regular convolution with depthwise separable convolution and $$1 \times 1 $$ 1 × 1 convolution. Moreover, the addition operation in the long skip connection was replaced with the concatenation operation to properly convey the information in the encoding phase. Besides, we propose the confusion block that can execute the model to proceed training more actively. From these approaches, we achieved a significantly small-sized network for fire segmentation with the highest performance. We compared the performance of the proposed method with various deep learning-based binary segmentation networks and image processing algorithm. Extensive experimental results on the FiSmo Dataset and Corsican Fire Database demonstrated that the proposed network outperforms other models with fewer parameters and is suitable for application in embedded devices.