Learning to segment objects of various sizes in VHR aerial images

Abstract

The goal of semantic segmentation is to assign semantic categories to each pixel in an image. In the context of aerial images, it is very important to yield dense labeling results, which can be applied for land use and land change detection. But small and large objects are difficult to be labeled correctly simultaneously in a single framework. Convolutional neural networks (CNN) can learn rich features and has achieved the state-of-the-art results in image labeling. We construct a novel CNN architecture: Pyramid Atrous Skip Deconvolution Network (PASDNet), which combines features of different levels and scales to learn small and large objects. Secondly, we employ a weighted loss function to overcome class imbalance problem, which improves the overall performance. Our proposed framework outperforms the other state-of-art methods on a public benchmark.