改进SSD模型在高分二号遥感影像中高速公路收费站位提取的应用

以高分二号遥感影像中的高速公路收费站为研究对象，选取了北京、山西、河南、广东、福建5个省市2019年的高速公路收费站点位和0.8 m遥感影像，通过图像预处理、样本标注、裁切、数据增强、样本集划分的步骤制作训练样本集；引入“多尺度特征融合”的方法对SSD目标检测模型进行改进，通过增加“转置卷积”和“拼接”操作，将高层次特征图像的语义特征赋予低层次特征图像，以增强上采样质量与特征融合能力，从而提升了模型对小目标收费站的检测效果；将改进SSD模型用于2019年福建省高分二号影像中的收费站点位提取，沿福建省高速公路路网矢量对影像进行自动切片，将切片输入模型中进行目标检测；保留有收费站的切片，使用非极大值抑制去除多余的检测框，将剩余的检测框的坐标变换为中心点的坐标，可以直接输出得到高速公路收费站的中心点矢量，从而实现对于收费站点位的端到端自动化提取。研究结果表明:改进SSD模型的精度、召回率及二者的调和平均数分别为0.86、0.88和0.87，均优于传统的SSD, VGG, Faster R-CNN和特征金字塔网络模型。可见，对收费站点位的自动提取可以大大提高公路管理者的工作效率，有效满足公路管理者的实际工作需求。 

Improved SSD model in extraction application of expressway toll station locations from GaoFen 2 remote sensing image

The locations of expressway toll stations from GaoFen 2 remote sensing images were extracted as the research object. Expressway toll stations and 0.8 m remote sensing images of Beijing, Shanxi, Henan, Guangdong and Fujian in 2019 were selected to create a training sample dataset via image preprocessing, sample labeling, cropping, data enhancement, and sample dataset partition. Multiscale feature fusion was introduced to improve the target detection model of the single-shot multibox detector (SSD) by adding two operations, namely, "deconvolution" and "concat." The semantic features of high-level feature maps were assigned to low-level feature maps to enhance the upsampling quality and feature fusion capabilities, thereby improved the detection performance on small targets toll stations. The improved SSD model was applied to extract the locations of toll stations in Fujian in 2019 from GaoFen 2 images. The images were automatically sliced along the Fujian highway network vectors, and the slices were input into the model for target detection. The slices with toll stations were retained, and non-maximum suppression was adopted to remove redundant detection frames. The coordinates of the remaining detection frames were transformed into the coordinates of the center points, and the center point vectors of the expressway toll stations were directly output. Thus, the automatic end-to-end extraction of toll station locations could be realized. Research results show that the accuracy and recall of the improved SSD model and their harmonic average are 0.86, 0.88, and 0.87, respectively, which are higher than those of the conventional SSD, VGG, Faster R-CNN, and Feature Pyramid Networks (FPN) models. Therefore, the proposed automatic extraction method for toll station locations can considerably improve management efficiency and adequately satisfy the actual needs of highway managers. 3 tabs, 7 figs, 35 refs.