改进U-Net网络及在遥感影像道路提取中的应用

高分辨率遥感图像分割在军事、民用等领域具有良好的应用前景，但由于复杂的背景条件以及干扰物的遮挡，导致现有算法无法较好地从遥感影像中提取道路细节信息。研究基于改进U Net网络模型，提出了MDAU-Net（multi dimension attention U-Net）网络结构模型，通过对U-Net网络结构加深至七层结构来提升精细分割道路的能力；并提出了一种多维注意力模块MD-MECA（multi dimension modified efficient channel attention），将其添加至编码部分的特征传递步骤中，以达到对编码部分的特征传递进行优化的目的；其中利用DropBlock与Batch Normalization解决网络训练过程中出现的过拟合。试验结果表明:改进后算法可以有效提升道路的提取效果，在测试集上的准确率达到了97.04%。

Improved U-Net network and its application of road extraction in remote sensing image

High resolution remote sensing image segmentation has a good application prospect in the military and civil fields, but due to the complex background conditions and the obstruction of interferences, the existing algorithms can’t extract road details from remote sensing images. Based on the improved U-Net network model, MDAU-Net (multi dimension attention U-Net) network structure model was proposed. The U-Net network structure was deepened to a seven-layer structure to improve the ability of fine segmentation of roads. A new multi dimensional attention module,which was called MD-MECA (multi dimension modified efficient channel attention), was proposed to optimize the feature transfer in the coding part. DropBlock and Batch Normalization were used to resolve the overfitting during network training. The experimental results show that the improved algorithm can effectively improve the road extraction effect, and the accuracy rate on the test set reaches 97.04%.