Multi-directional Local-Global feature Fusion Network for Building Segmentation in Remote Sensing Images

High-resolution remote sensing images can provide important information for building entity segmentation. However, the arbitrariness of building orientation makes it difficult to automatically segment buildings with high accuracy. Current algorithms mostly use horizontal strip convolution and vertical stripe convolution in the building extraction task, which is not sufficient to extract accurate boundary positioning information and complete footprints, especially in high-resolution images. In this article, we propose a novel Multi-directional Local-Global feature Fusion Network (MLGF-Net) based on multi-directional strip convolution. To better obtain edge localization information, we designed a multi-directional strip convolution joint module (MDSJ). This module extends the directionality of the extracted features to 0°, 45°, 90°, and 135° by parallel connection fusion so that the feature extraction capability of the model is enhanced. In addition, to improve the completeness of the building footprint, we proposed the local-global double-branch structure (LGS) module. These two branches are structured in parallel and utilize the self-attention mechanism to obtain more local-global information about the building. Extensive experiments on the Massachusetts dataset demonstrate the superior performance of the proposed method.