Authors: Jialin Li*, The Ohio State University, Ningchuan Xiao, The Ohio State University
Topics: Geographic Information Science and Systems, Cartography
Keywords: deep learning, cartography, map element detection, choropleth map
Session Type: Paper
Start / End Time: 8:00 AM / 9:15 AM
Room: Governors Square 10, Sheraton, Concourse Level
Presentation File: No File Uploaded
Automatically understanding a map image by machines is a challenge in artificial intelligence era. This paper aims at automatically detecting map elements in choropleth maps, which can be further utilized to understand maps in general. Specifically, if we want to know the theme of a choropleth map (e.g. adult obesity rate map), we need to first find the title of the map and then conduct semantic analysis on the title. Deep learning based object detection methods, including Faster Region-based Convolutional Neural Network (Faster R-CNN) and You Only Look Once (YOLO), will be applied to detecting map titles and legend areas of choropleth maps. For both detectors, the input data is choropleth map images, and the output is locations and sizes of bounding boxes in the images and the corresponding element classes. More than 1000 choropleth map images with either titles or legends will be collected using Google Images or from other sources. And the images will be used to train the detection models. Each of the two methods will be evaluated by mean Average Precision (mAP) and frame rate (number of images processed per second) for detection accuracy and speed respectively. This study intends to demonstrate the ability of Faster R-CNN and YOLO in detecting elements of choropleth maps, and comparisons of the detection accuracy and speed of the methods will be made to understand the strength of each method. This research will also provide suggestions about how to select detectors in different conditions for other map element detection tasks.