Fire Detection using Residual Deeplabv3+ Model

Fire detection and prevention had become a high priority task, in the last decade, due to the higher number of forest fires. Automatic detection systems facilitate the intervention and reduce the cost of firefighters travel in case of false occurrences. Deep learning based systems has drawn promising high results in the field, in particular, Deeplabv3+ which is an architecture based on the so called Atrous Spatial Pyramid Pooling that enhances the segmentation results. This paper present the results of Deeplabv3+ applied over the french Corsican dataset with an Xception backbone. The model had been trained over the RGB collection of pictures of the dataset and over the whole dataset that englobes RGB and infrared (IR) pictures. Several experiments using Dice and Tversky loss functions are conducted in order to further reduce the problems induced by unbalanced datasets. The performance is measured with some extended metrics namely: mean Intersection over Union (IoU), mean Boundary F1 (BF) Score, and mean Dice similarity measure in addition to the standard accuracy metric. The achieved results demonstrate that the Deeplabv3+ with Xception gives very encouraging results for fire detection over RGB and IR images.