非对称卷积金字塔残差网络的遥感影像建筑物提取  

Building Extraction from Remote Rensing Images Using Asymmetric Convolution Pyramid Residual Network

作　　者：吴锋振 杨德宏[1] 李俊[1] 胡明洪 WU Fengzhen;YANG Dehong;LI Jun;HU Minghong(School of Land and Resources Engineering,Kunming University of Science and Technology,Kunming 650093,China)

机构地区：[1]昆明理工大学国土资源工程学院,云南昆明650093

出　　处：《遥感技术与应用》2023年第6期1467-1476,共10页Remote Sensing Technology and Application

摘　　要：针对建筑物提取中存在微小建筑物提取不完整、建筑物边界提取不准确等问题,以ResNet为基础,将残差模块与Inception-V3模块嵌套,然后进行金字塔池化来改善网络分割性能,并将改进的神经网络模型称为非对称卷积金字塔残差网络(In_PPM_ResNet)。该模型兼具三者的优势:残差模块能解决网络退化问题;Inception-V3模块中使用非对称卷积,减少参数量,节约运算资源,多路径级联增加网络宽度;金字塔池化模块获取不同尺度信息,扮演加强特征提取角色。为验证In_PPM_ResNet的有效性和适用性,分别在WHU和AIRS建筑物数据集上进行实验,并将其与典型建筑物提取网络FCN-8S、UNet++、SegNet进行对比。实验结果表明:在WHU数据集上,In_PPM_ResNet相较于其它网络,IoU为最优值,达89.97%,其他评价指标也有一定程度提升,模型参数量对比网络中排名第3,参数量相对较小,每轮次运行时间与其他网络相差不大,证明该模型的效率相对较高;在0.075 m分辨率的AIRS数据集上,交并比、召回率、F1分数均为最优值,分别达88.52%、88.95%和87.85%。另一方面,建筑物提取结果相比于其他网络,边界更准确,空洞少,有一定的应用潜力。Incomplete extraction of tiny buildings and inaccurate boundaries are two major problems when ex⁃tracting buildings from high-resolution remote sensing images by a convolution neural network.To solve these problems,we proposed a building extraction method to improve the accuracy of building extraction.Method on the basis of deep residual network(ResNet),we proposed a building extraction network by nesting residual mod⁃ule and Inception-v3 module and carrying out pyramid pooling.The network is referred to as asymmetric convo⁃lution pyramid residual network(In_PPM_ResNet).The model combines the advantages of all three:the residu⁃al module can solve the network degradation problem;the use of asymmetric convolution in the Inception-V3 module reduces the number of parameters and saves computing resources,and the multi-path cascade increases the network width;the pyramid pooling module acquires information at different scales and plays an enhanced feature extraction role.To verify the effectiveness and applicability of In_PPM_ResNet,experiments are con⁃ducted on WHU and AIRS building datasets respectively,and compared with typical building extraction net⁃works FCN-8S,UNet++,and SegNet.The experimental results show that,on the WHU dataset,In_PPM_ResNet has the optimal value of IoU of 89.97%compared with other networks,and other evaluation indexes have also been improved to some extent,the number of parameters of the model ranks the third in the comparison network,and the number of parameters is relatively small.The running time of each round is not dif⁃ferent from that of other networks,which proves that the efficiency of the model in this paper is relatively high.On the 0.075m resolution AIRS dataset,IoU,Recall,and F1-score are all optimal,reaching 88.52%,88.95%,and 87.85%,respectively.On the other hand,the building extraction results have more accurate boundaries and fewer holes compared with other networks,which have certain application potential.

关 键 词：建筑物提取 ResNet Inception-V3 非对称卷积 金字塔池化模块 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]