R-DoGAN: Task-Specific Convolutional Neural Network for Rock Fracture Segmentation Using Perceptual Loss and Edge Information Input

抄録

We introduce a task-specific convolutional neural network for rock fracture image segmentation, named R-DoGAN (Rock-DoG-GAN), which integrates distinct features of fracture segmentation into the network training and input processes. Instead of using low-dimensional information for network training, a generative adversarial network (GAN)-based perceptual loss is used; and the difference of Gaussians (DoG) images, which contain multi-resolution edge information, are used as additional network input. Test results demonstrate that R-DoGAN outperforms previous networks, despite having fewer network parameters and a smaller training dataset.