Deep Neural Network with Walsh-Hadamard Transform Layer For Ember Detection during a Wildfire

Hongyi Pan, Diaa Badawi, Chang Chen, Adam C. Watts, Erdem Koyuncu, Ahmet Enis Çetin

Abstract: In this article, we describe an ember detection method in infrared (IR) video. Embers, also called firebrands, can act as wildfire super-spreaders. We develop a novel neural network with a Walsh-Hadamard Transform (WHT) layer to process the IR video. The WHT layer is used to process the temporal dimension of the video data to model the high-frequency activity due to ember movements. We insert the WHT layer to ResNet-18 and obtained higher accuracy compared to the standard single slice ResNet-18 and the ResNet-18 processing the entire video block. We also repeat the experiments on ResNet-34, but we found that ResNet-18 is sufficient for this task. Therefore, we choose the ResNet-18 with the WHT layer as the proposed model.

External IDs:dblp:conf/cvpr/PanBCWKC22