基于改进U-Net的遥感图像道路检测

Road detection in remote sensing images based on improved U-Net

摘要:

针对复杂遥感影像中因道路材质多样、背景复杂以及阴影遮挡等因素所导致的识别精度不高，以及现有模型参数量大等问题，设计了一种基于U-Net的改进模型，用于遥感图像道路检测；首先，在编码器中引入MobileNetV2代替原有的主干提取网路，在保证检测精度的同时大幅度降低计算复杂度；其次，在解码器部分我们引入了深度可分离卷积，该方法将空间特征提取与通道融合拆分处理，以更少的运算量完成对高分辨率路面细节的恢复；最后，引入空间组增强注意力机制，强化模型对空间层级特征的建模能力，从而提升对目标区域的感知精度与检测效果；实验结果表明，相较于对照模型，本文改进模型在马萨诸塞州道路数据集上的mIOU和F1-score分别达到了最高的79.50%和87.46%；实现了检测性能与模型复杂度之间的有效平衡，为遥感影像中的道路提取提供了有力支持。

Abstract:

Aiming at the low recognition accuracy in complex remote sensing images due to the variety of road materials, complex background and shadow occlusion, as well as the large number of parameters in the existing models, this paper designs an improved model based on U-Net for road detection in remote sensing images. Firstly, MobileNetV2 is introduced into the encoder instead of the original backbone extraction network, which ensures the detection accuracy while greatly reducing the computational complexity. Secondly, in the decoder part, depthwise separable convolution is introduced to separate spatial feature extraction and channel fusion processing, thereby accomplishing the recovery of high-resolution road details with reduced computation. Finally, the spatial group enhancement attention mechanism is introduced to strengthen the model"s ability to model spatial level features, so as to improve the perception accuracy and detection effect on the target area. The experimental results show that compared with the control model, the improved model in this paper achieves the highest mIOU and F1-score of 79.50% and 87.46% on the Massachusetts Roads Dataset, respectively. An effective balance between detection performance and model complexity is realized, which provides strong support for road extraction in remote sensing images.