Data augmentation using image-to-image translation for detecting forest strip roads based on deep learning

摘要

Having been developed recently, image classification and object detection by deep convolutional neural networks are now widely used. However, in applications of deep learning in forestry, hardly any cases have involved forestry robots. For the autonomous driving and working of a forwarder on a strip road, a system is developed for detecting strip roads by semantic segmentation using deep learning, and data augmentation methods are proposed on the basis of generative adversarial networks (GANs) to improve robustness. In this study, three GAN-based data augmentation methods are proposed, namely, (i) translated images from new label images, (ii) translated images from an actual dataset, and (iii) both. The training dataset is evaluated by fully convolutional networks, from which the trained models show a pixel accuracy of 0.616 and a mean accuracy of 0.512. Compared with no augmentation and general augmentation, a maximum improvement in accuracy of 0.031 is observed. The GAN-based augmentation technique is effective for detecting a small number class because the class distribution of the dataset is set arbitrarily. Accurate detection by the trained model is confirmed even if the image dataset contains unknown obstacles.