基于改进的DUNet遥感图像道路提取

Road extraction from remote sensing images based on improved DUNet

为进一步提高遥感图像道路提取的精度,提出一种改进的DUNet遥感图像道路提取方法。在编码器部分,为使网络关注道路信息,在第3个池化层分别使用有注意力机制和没有注意力机制两个分支提取道路特征;在解码器部分,同时使用传统UNet的解码器和DUNet解码器两个分支进行上采样,最大限度减少信息丢失。试验结果表明,与其他8种常用的分割模型结果相比,此方法在Massachusetts和DeepGlobe 2018数据集上都获得最高的平均交并比和平均Dice系数,其中平均交并比最高分别提高2.90%和8.99%,平均Dice系数最高分别提高2.53%和7.66%。这表明改进的DUNet能够有效实现遥感图像的道路提取,与传统DUNet相比,在小路区域的分割效果得到提升,进一步提高了传统DUNet的分割精度。

In order to further improve the accuracy of road extraction from remote sensing images,an improved DUNet remote sensing image road extraction method was proposed.In the encoder part,to make the network pay attention to road information,two branches with and without attention mechanism were used to extract road features in the third pooling layer respectively.In the decoder part,upsampling was carried out by two branches of both the traditional UNet decoder and DUNet decoder,which minimized the loss of information.Experiment showed that the results were compared with those of eight other commonly segmentation models,this method achieved the highest mean intersection over union and mean dice coefficient on both the Massachusetts and DeepGlobe 2018 datasets,among which,mean intersection over union increased by 2.90%and 8.99%,and mean dice coefficient increased by 2.53%and 7.66%.The improved DUNet could effectively realize road extraction from remote sensing images.Compared with traditional DUNet,the segmentation effect in the path area was improved,and the segmentation accuracy of traditional DUNet was further improved.