Abstract: Aiming at the wrong segmentation caused by the large scale change of buildings, multiple confounding and occlusion in high-resolution remote sensing images, a building extraction method based on multi-scale MLP and edge thinning is proposed. Firstly, the local multilayer perceptron and the global multilayer perceptron are combined to improve the model’s ability to understand and perceive different areas in the image, and weaken the interference of background independent elements. Then, by obtaining edge enhancement images of different level feature maps, the refinement of building contours is achieved. Finally, a feature level deep supervision module is designed to balance the positioning accuracy and edge accuracy of the output prediction map, further improving the segmentation effect. The experimental results show that the proposed algorithm achieves 96.18%, 95.74%, 95.96%, 91.82% and 91.37%, 89.66%, 90.51%, and 82.79% precicion, recall, F1 score, and IoU, respectively, on the WHU and Inria Aerial Image Labeling datasets. Compared to other related algorithms, it not only improves accuracy to varying degrees but also maintains a lower parameter volume and computational load, providing robust support for the accurate and rapid extraction of building information from optical high-resolution remote sensing images.

Key words: multilayer perceptron, building segmentation, multi-scale MLP, edge thinning

摘要： 针对高分辨率遥感影像中建筑物尺度变化大、干扰因素多和被遮挡等问题导致的错误分割，提出一种融合多尺度MLP和边缘细化的遥感影像建筑物提取方法。结合局部多层感知器与全局多层感知器提高模型对于影像中不同区域的理解能力和感知能力，弱化背景无关元素的干扰；通过获取不同层次特征图的边缘增强图，实现建筑物轮廓的精细化；设计一个对输出预测图进行特征级深度监督模块来兼顾定位精度和边缘精度，进一步提高分割效果。实验结果表明，所提算法在WHU和Inria Aerial Image Labeling数据集上的precision、recall、F1-score和IoU分别达到了96.18%、95.74%、95.96%、91.82%与91.37%、89.66%、90.51%、82.79%，对比其他相关算法，精度在不同程度提升的同时也保持了较低的参数量和计算量，为准确、快速地提取光学高分辨率遥感影像中的建筑物信息提供了有力支持。

关键词: 多层感知器, 建筑物分割, 多尺度MLP, 边缘细化