融合双域特征均衡的遥感图像道路提取

当前遥感图像道路提取模型仍在很大程度上受道路植被遮挡所影响，导致网络模型对道路信息误判。为此，基于双域特征均衡提出了一种不受遮挡物影响的道路提取方法，高效地实现植被遮挡下的道路提取。具体而言，提出了一种新的道路提取卷积神经网络，该网络由去除遮挡子网络和道路提取子网络组成。在去除遮挡子网络中嵌入一个分层卷积模块用于提取输入图像的深层结构特征和浅层纹理特征，以及双域均衡模块用于特征还原，以此去除目标道路上的遮挡物。道路提取子网络用于对去除遮挡后的道路结构进行精细的分割，得到准确性更高的道路提取结果。通过在四川西南农村地区的遥感数据集上进行大量实验，结果显示基于双域特征均衡的方法相较于其他遥感图像道路提取方法在像素精确度(Overall Accuracy,OA)和交并比(Intersection over Union，IoU)指标上达到了最高，分别是98.16%和85.38%。

Remote Sensing Image Road Extraction Combined with Two-omain Feature Equalization

The existing remote-ensing image road extraction techniques still produce misclassification for roads obscured by occluders.To this end,a road extraction method based on two-omain feature equalization that is not affected by occluders is proposed,which can efficiently achieve road extraction under vegetation occlusions.Specifically, a new convolutional neural network（CNN） for road extraction is proposed,which consists of an occlusion removal subnetwork and a road extraction subnetwork.The former embeds a hierarchical convolution module to extract the input image- deep structural features and shallow texture features.In addition,the subnetwork includes a dual-omain equalization module for feature restoration to remove occluders from the road.The latter sub-etwork uses a mainstream segmentation network to accurately extract the roads after removing the occluders to obtain more accurate road extraction results.Through extensive experiments in the rural areas of southwest Sichuan Province,the results show that the proposed method achieves the highest overall accuracy (OA) and intersection over union (IoU) indexes compared with other remote-ensing image road extraction methods,which are 98.16