Abstract:

The extraction of road targets from remote sensing images is of great significance to smart city construction. Due to the complexity and diversity of road and background features in remote sensing data, the accuracy of road extraction using deep learning methods is limited. Therefore, based on the U-Net network architecture design, a deep semantic segmentation model AS-Unet for road extraction from remote sensing images is implemented. The model is divided into two parts：encoder and decoder. The algorithm first adds a channel attention mechanism to the encoder, to filter the extracted rich low-level features, highlight target features, inhibit background noise interference, and improve the accuracy of deep and shallow information fusion. Secondly, considering the single size sensitivity issue of the network to the road targets in the images, a spatial pyramid pooling module is added after the last convolutional layer of the encoding network, to capture road features of different scales. Finally, a spatial attention mechanism is added to the decoder, to further perform learning of location relationship information, and filtering of relevant deep semantic feature, and to improve the ability of feature map restoration. Experiments are conducted on Massachusetts and DeepGlobe road datasets, and the results demonstrate that compared to semantic segmentation networks such as SegNet and FCN, the AS-Unet network is preferable in terms of the evaluation indexes such as recall, precision, and [F1] value. The designed AS-Unet network has satisfactory performance and higher segmentation accuracy, and boasts certain theoretical and practical application value.

Key words: semantic segmentation, high-resolution image, road extraction, attention mechanism, spatial pyramid model, convolutional neural network

摘要：

从遥感影像中提取道路目标对智慧城市建设具有重要意义。由于遥感数据中道路及背景特征复杂多样，使用深度学习方法对道路进行提取的准确性仍然受到限制。基于U-Net网络架构设计实现了用于遥感影像道路提取的深度语义分割模型AS-Unet，该模型分为编码器和解码器两部分。在编码器部分加入通道注意力机制，对提取的丰富低层特征进行筛选，突出目标特征，抑制背景噪声干扰，从而提高深浅层信息融合准确率；为解决网络对道路目标单一尺寸的敏感问题，在编码器最后一层卷积层后面加入空间金字塔池化模块来捕获不同尺度道路特征；在解码器部分加入空间注意力机制，进行位置关系信息学习和深层次语义特征筛选，提高特征图还原能力。在Massachusetts和DeepGlobe道路数据集上进行实验，结果证明，在召回率、精度、[F1]值等评估指标上，明显优于SegNet、FCN等语义分割网络。所设计的AS-Unet网络性能优良，具有更高的分割准确率，具备一定理论和实际应用价值。

关键词: 语义分割, 高分辨率影像, 道路提取, 注意力机制, 空间金字塔模型, 卷积神经网络