Holistically guided feature for semantic segmentation of high resolution remote sensing images

Abstract

Semantic segmentation plays a crucial role in facilitating human understanding of remote sensing images. In recent years, significant advancements have been made in remote sensing semantic segmentation, largely attributed to the remarkable success of the fully convolutional network (FCN) in computer vision. The encode-decode architecture leverages the encoder to produce a continuous flow of semantic information, which are seamlessly integrated within the decoder. To fully exploit the potential of semantic features, several attention modules have been introduced to incorporate intermediate contextual information. However, these complex models often demand substantial computational resources. To tackle this challenge, this study presents a sophisticated model called the holistically guided fully convolution network (HGFCN). Unlike the conventional U-Net approach, which connects each relevant layer of the encoder and decoder, our model leverages holistically guided features derived from middle and high-level semantic features to recover spatial information. Empirical evaluations conducted on two high-resolution remote sensing datasets substantiate that the proposed method surpasses state-of-the-art semantic segmentation models.