Feature Pyramid Network Embedded with Encoder-Decoder Architecture for Semantic Segmentation of High-Resolution Aerial Imagery

Abstract

Semantic segmentation of high-resolution aerial imagery is a crucial task in remote sensing, which enables the pixel-level understanding of complex peri-urban and urban environments. Accurate segmentation remains challenging due to class imbalance, fine boundary ambiguity, and presence of multi-scale objects. In this paper, a comprehensive feature pyramid network (FPN) combined with encoder-decoder-based semantic segmentation architectures is evaluated on a publicly available highresolution aerial imagery dataset. The dataset is pre-processed using a patch extraction technique, followed by data augmentation to optimize model performance. Furthermore, we experimented with different segmentation frameworks and, based on the evaluation metrics, the ResNeXt-101 backbone in combination with the FPN framework outperformed other models for this multi-resolution semantic segmentation problem. 

Mehr zum Titel