文献类型： 外文期刊

作者： Dong, Luan;Dong, Luan;Zheng, Fengling;Yan, Qin;Zheng, Fengling;Dong, Luan;Dong, Luan

作者机构：

关键词： Topographic maps; Graticule intersections; Object detection; Convolutional neural networks; Geometric calibration

期刊名称： MACHINE VISION AND APPLICATIONS

ISSN： 0932-8092

年卷期： 2019 年 30 卷 4 期

页码：

摘要： Graticule intersections in topographic maps are usually considered to be suitable candidates for reference points in geometric calibration because the corresponding geographical information can be directly retrieved from the maps or derived from sheet numbers. Previous research on automatic corner point detection relies on the assumption that scanned maps are not rotated, which is rarely practical. To address this issue, a semantic segmentation approach for accurate graticule intersection localization is proposed in this paper. A fully convolutional network is utilized to provide pixel level information about the locations of specific rectangular objects at the corners of map frames by dense classification in regions of interest. The globally optimal segmentation of the foreground rotated object is obtained by the graph cuts technique. The bounding box of the rotated object is further retrieved with the minimum-area enclosing rectangle algorithm. Finally, the coordinates of graticule intersections are derived in accordance with the positions of the sliding windows and the relative locations of the vertices of the objects. The proposed method reduces the average localization error to 1.5 pixels, which is 32.4% lower than that of the baseline model. The standard deviation of localization error is 0.91 pixels, which aligns with an average of 52% improvements to the baseline model in the location variance metric.

分类号：