面向自动驾驶的遥感影像路网检测方法

针对高分辨率遥感影像道路提取难度大、提取精度低的问题,提出了一种基于VGGU-Net++的遥感影像路网检测方法。首先基于VGGU-Net框架构建了编码器-解码器网络;其次,设计了一系列嵌套、密集的卷积块,用以缩小编码器与解码器特征映射之间的语义差距。节点之间利用跳跃连接填充了具有多个卷积层的密集卷积块,其层数取决于金字塔等级;并在2个卷积块之间设置1个串联层,该层将同一密集块前一个卷积层的输出和浅层的上采样输出进行特征图融合。同时,使用深监督策略保证网络模型的修剪程度和速度增益。在网络训练过程中,这种相似语义特征图的跳转连接可以简化优化操作,提高网络训练性能。最后,利用遥感影像开源数据集——马萨诸塞州数据集进行算法的测试与验证。结果表明,提出的VGGU-Net++网络与现有同类方法相比,获得了更好的性能表现,在精确率、召回率、F₁-score和I_oU方面分别达到了88.1%、87.1%、88.5%和77.9%,能够实现城市、山区、直线、弯曲道路场景高维、复杂、抽象特征的自动提取;此外,检测结果能够减少干扰,降低误检,保留更多道路细节信息,得到更加完整、清晰的路网检测效果。

Road Network Detection of Remote Sensing Images for Autonomous Driving

To address the difficulty and low accuracy of road extraction from high-resolution remote sensing images, this study presents a road network detection method based on VGGU-Net++. First, an encoder-decoder framework is constructed based on the VGGU-Net network. Second, a series of nested and dense convolution blocks is designed to narrow the semantic gap between the feature mapping of the encoder and decoder. Subsequently, a dense convolution block with multiple convolution layers is filled with skip connections between nodes. The number of layers depends on the pyramid level, and a concatenation layer is set between the two convolution blocks, which fuses the feature map of the output of the previous convolution layer of the same dense block with the up-sampling output of the shallow layer. In addition, a deep supervision strategy guarantees the pruning degree and speed gain of the network model. Finally, the proposed method is tested and verified using an opensource dataset (the Massachusetts dataset) of remote sensing images. The results show that the VGGU-Net++ network achieves a better performance than existing deep learning methods, achieving 88.1%, 87.1%, 88.5%, and 77.9% in precision, recall, F₁-score, and I_oU, respectively. It can automatically extract high-dimensional, complex, and abstract features of urban, mountainous, straight, and curved road scenes. The detection results can reduce interference and false detections, retain more road details, and obtain clearer road-network detection results.