针对在遥感影像建筑物提取过程中,建筑物与周围环境信息混淆导致提取精度下降的问题,提出了一种低维特征信息增强的改进U型卷积神经网络(U-net)模型,用于遥感影像建筑物的提取。借鉴医学影像分割中应用广泛的U-net模型对建筑物进行提取;考虑到在网络传播过程中低维细节信息逐级削弱,在特征金字塔中的特征图与扩张路径同级上的特征融合前,先与上一层级的特征图进行融合,进一步优化了提取结果的边缘提取精度。在覆盖范围约340 km 2的遥感影像数据集上进行实验,结果表明本文提出的方法在交并比、像素精度和Kappa系数3个指标上的均值分别达到83.9%、92.8%和83.6%,均优于模糊C均值、全卷积网络与经典U-net方法。

The building environment in urban areas is complex. Achieving high building detection accuracy from remote sensing images is challenging because of the difficulty associated with distinguishing between buildings and the environmental information. To solve this problem, an improved U-type convolutional neural (U-net) network with enhanced low-dimensional feature information is proposed for detecting buildings from the remote sensing images. Initially, a building is detected using the U-net network model typically employed for medical image segmentation. Further, the low-dimensional information is weakened at each step of the network propagation process. Before merging the feature map of a certain level in the feature pyramid with the feature map of the corresponding expansion path level, it is merged with the feature map of the previous level to optimize the detection accuracy of the building edges. According to the experimental results obtained using a dataset of remote sensing images covering a range of approximately 340 km ², the proposed method achieves values of 83.9%, 92.8%, and 83.6% for the intersection-over-union (IoU), pixel accuracy, and Kappa coefficient, respectively, demonstrating its superior performance when compared with the fuzzy C-means clustering algorithm, fully convolutional neural network, and classic U-net methods.