Pre-processing for UAV Based Wildfire Detection: A Loss U-net Enhanced GAN for Image Restoration

Abstract

In this paper, a U-net with feature loss enhanced generative adversarial network (GAN) is designed for the wildfire or smoke images restoration which is captured by unmanned aerial vehicles in a serious environment. Based on the concepts of GAN, feature loss, and fastai API, we firstly crappy the target images, and train a U-net architecture based generator, then load the adaptive loss of discriminator and the mean square error together to train the GAN model. After the GAN, a second U-net grabs the feature loss from an Imagenet pre-trained loss network to generate the GAN output images with one more step. This U-net enhanced the generator of GAN and helped to get the main features in human conception. Comparing with other restoration methods, this model used the adaptive loss to train the GAN and perceptual loss to train the next U-net. Learning rate with simulation annealing helped jumping out of the local minimum. The result proved the good performance of this model.