高分辨率遥感影像建筑物提取卷积神经网络

Convolutional neural networks for building extraction from high-resolution remote sensing images

针对复杂影像场景下卷积神经网络对建筑物提取效果较差的问题,本文对轻量型卷积神经网络LinkNet进行了优化调整,使用融合坐标注意力机制的深层残差网络CA-ResNet-50作为LinkNet的编码器,显著增强了网络模型的特征提取性能,提升了其在复杂场景下建筑物提取能力;同时,利用卷积分解方法对LinkNet初始块进行优化,获得了更快的网络训练速度,最终得到性能优异的建筑物提取网络CA-LinkNet。试验结果表明,在武汉大学航空建筑物数据集上CA-LinkNet与最初的LinkNet相比精度指标IoU、Kappa和F1分别提升了2.01%、1.26%和1.11%。此外,在选取的数据集上CA-LinkNet各项精度指标均优于经典分割网络,在复杂影像场景下也能有效提取建筑物,表现出较强的抗干扰能力。

Convolutional neural networks have poor effects on building extraction in complex image scenes.To address this issue,this paper optimized and adjusted the lightweight convolutional neural network LinkNet.CA-ResNet-50,a deep residual network integrated with coordinate attention(CA)mechanism,was used as the encoder of LinkNet,which significantly enhanced the feature extraction performance of the network model and improved its building extraction capability in complex scenes.At the same time,the convolutional decomposition method was used to optimize the initial block of LinkNet,and a faster network training speed was obtained.Finally,the building extraction network with excellent performance CA-LinkNet was obtained.The test results show that the accuracy indexes IoU,Kappa,and F1 of CA-LinkNet on the WHU aerial building dataset have increased by 2.01%,1.26%,and 1.11%,respectively,compared with the original LinkNet.In addition,the accuracy indexes of CA-LinkNet on the selected dataset are superior to those of the classical segmentation network,and CA-LinkNet can effectively extract buildings in complex image scenes,showing strong antiinterference ability.