基于多任务学习的高分辨率遥感影像建筑实例分割

针对基于深度神经网络的高分辨率遥感影像建筑物提取算法中将建筑物提取视为二分类问题(即将遥感影像中的像素点划分为建筑物与非建筑两类)而无法区分建筑物个体的局限性, 将基于Xception module改进的U-Net深度神经网络方法与多任务学习方法相结合进行建筑物实例分割, 在获取建筑物二分类结果的同时, 区分不同建筑物个体, 并选择Inria航空影像数据集对该方法进行验证。结果表明, 在高分辨率遥感影像的建筑物二分类提取方面, 基于Xception module改进的U-Net方法明显优于U-Net方法, 提取精度升高1.4%; 结合多任务学习的深度神经网络方法不仅能够实现建筑物的实例分割, 而且可将二分类建筑物的提取精度提升约0.5%。

Instance Segmentation of Buildings from High-Resolution Remote Sensing Images with Multitask Learning

At present, building extraction from high-resolution remote sensing images using deep neural network is viewed as a binary classification problem, which divides the pixels into two categories, building and nonbuilding, but it cannot distinguish individual buildings. To solve this problem, the U-Net modified with Xception module and multitask learning are combined to apply to the instance segmentation of buildings, which both acquires the binary classification and distinguishes the individual buildings. Inria aerial imagery is used as the research dataset to validate the algorithm. The results show that the binary classification performance of U-Net modified with Xception outperforms U-Net by about 1.4%. The multitask driven deep neural network not only accomplishes the instance segmentation of buildings, but also improves the accuracy by about 0.5%.