级联融合边缘特征的高分辨率遥感影像道路提取

Road Extraction From High-Resolution Remote Sensing Images Based on Cascade Fusion of Edge Features

针对进行高分辨率遥感影像道路提取时常出现的识别错误和提取结果断裂等问题,提出一种级联融合边缘特征和语义特征的ACEResUNet多任务融合模型。该模型通过边缘检测任务进行道路边缘特征自动化提取,将其与改进的ResUNet模型对应的卷积单元进行特征级联融合,为语义分割道路训练提供更多的决策依据,提升道路提取结果的连通性。通过在各模型特征提取单元中引入交叉压缩注意力模块,提升模型的特征提取能力,并在改进的ResUNet模型的编解码器之间添加全局多尺度特征融合模块,获取不同尺度目标地物的全文特征信息,以提升道路最终提取结果的完整性。在DeepGlobe道路数据集上的实验结果表明,该模型的道路提取精确率和交并比分别达到了0.798和0.661,相较于VNet和ResUNet等经典模型均有提升。

Aiming at the problems of recognition errors and broken extraction results that often occur in road extraction from high-resolution remote sensing images,this paper proposes a multi-task fusion model of ACEResUNet that cascades edge features and semantic features.The model automatically extracts road edge features through the edge detection task,and fuses it with the convolution unit corresponding to the improved ResUNet model for feature cascade fusion,which provides more decision-making basis for semantic segmentation training and helps to accurately segment the road boundary.By introducing a cross-compression attention module in each model feature extraction unit,the feature extraction capability of the model is improved,and a global multi-scale feature fusion module is added between the encoders and decoders of the improved ResUNet model to obtain the target features of different scales.Also full-text feature information is obtained to improve the integrity of the final road extraction results.The experimental results on the DeepGlobe road dataset show that the road extraction accuracy and intersection ratio of this model reach 0.798 and 0.661,respectively,which are greatly improved compared to classic models such as VNet and ResUNet.