| Popis: |
Semantic segmentation is a basic task in the interpretation of remote sensing images. Mainstream deep-learning-based semantic segmentation algorithms typically process images with small sizes. However, remote sensing images typically involve large areas with buildings and water, which have weak textures. Because of the limited range of receptive fields, the semantic segmentation of such areas from small images may lead to problems, such as loss of spatial features and inaccurate boundary extraction. To address these problems, this article devises a patch-to-region framework for the semantic segmentation of large-format remote sensing images. This framework has a bottom–up pyramid structure, where the bottom layer is a small image patch, referred to as a “patch,” and the upper layer is a large image region, referred to as a “region.” The classical semantic segmentation network is first used to process small image patches to obtain pixel-by-pixel semantic features. Then, the pixel-by-pixel semantic features are sparsely reduced into patch-level semantic feature vectors, and the semantic feature vectors of the entire image region are processed through the contextual information extractor to extract the global semantic feature vectors. Subsequently, an information aggregation module is used to integrate the global semantic feature vectors and semantic features to obtain new semantic features with both global and local information. Finally, a lightweight decoding module is used to process the new semantic features to obtain the final semantic segmentation result. The developed framework is evaluated over three public datasets. The results of extensive experiments show that the framework can achieve more accurate and reliable semantic segmentation of high-resolution remote sensing images than state-of-the-art semantic segmentation algorithms. Moreover, ablation studies are performed to verify the effectiveness of each module of the framework. |
