A Generic Neighborhood Information Extraction Mechanism for High-resolution Remote Sensing Image Segmentation

Remote sensing images are usually too big to be segmented at once by convolutional neural networks (CNNs). Therefore, they need to be cut into smaller image patches for training, validation, and testing, which can lead to objects being cut off, often to the point of unrecognizability. As unrecognizable objects are hard to segment, the segmentation quality of the results suffers. This thesis presents a generic neighborhood information extraction (NIE) mechanism that allows segmentation architectures to access information from surrounding image patches of an input patch with the aim of restoring cut-off objects to increase the segmentation quality. With only a quarter of the baseline’s parameters, the best NIE architecture was able to achieve 91.5% of the baseline’s accuracy, and a feature map analysis of the NIE mechanism showed that it extracts segmentation-relevant information. Both of these findings underline the mechanism’s potential and offer promising future research directions.