Multiple object extraction from aerial imagery with convolutional neural networks

An automatic system to extract terrestrial objects from aerial imagery has many applications in a wide range of areas. However, in general, this task has been performed by human experts manually, so that it is very costly and time consuming. There have been many attempts at automating this task, but many of the existing works are based on class-specific features and classifiers. In this article, the authors propose a convolutional neural network (CNN)-based building and road extraction system. This takes raw pixel values in aerial imagery as input and outputs predicted three-channel label images (building-road-background). Using CNNs, both feature extractors and classifiers are automatically constructed. The authors propose a new technique to train a single CNN efficiently for extracting multiple kinds of objects simultaneously. Finally, they show that the proposed technique improves the prediction performance and surpasses state-of-the-art results tested on a publicly available aerial imagery dataset.