Thematic Map Automatic Reading Method Based On Faster R-CNN

Thematic maps provide a prominent and complete representation of the relevant content of a topic.They often have specialized uses.Realizing the automatic reading of thematic maps by computer to obtain information is conducive to improving the efficiency of map reading and giving full play to the value of the map.In this paper,object detection model is introduced into the task of automatic reading and information acquisition of thematic maps for the first time.Based on the bounding boxes of elements such as the zoning statistical diagram and the area symbols of administrative division units output by model detection,a method for computer to automatically read and obtain information and generate natural language descriptions from the content represented by regional classified statistic graph method and cartodiagram method in thematic map is proposed.Experiments verify the effectiveness of the method.And the transformation from map to text is realized by generating description text based on the template.Specifically,the research content and corresponding results of this paper mainly include the following points:(1)This paper uses the JPG file of the "Shenzhen Atlas" as the data source to produce thematic map data set for the map element detection of the cartodiagram method and regional classified statistic graph method.And a variety of enhancement methods are used to increase the size of the data set.(2)This paper uses the Paddle Paddle deep learning platform to build the Faster RCNN model based on the results of the model comparison experiment to perform the object detection experiment of the map elements of the cartodiagram method and regional classified statistic graph method.The highest m AP values obtained on the validation set of the object detection model of the cartodiagram method and regional classified statistic graph method are 92.590% and 98.542% after a series of optimization work including backbone network optimization,fusion of multi-scale features based on feature pyramid network,setting Anchor in consideration of feature attributes,optimization of regression loss function,optimization of non-maximum suppression methods,and post-detection processing.(3)This paper proposes a method to determine the filling color of the area symbols based on the detection of the bounding box of the administrative division unit,and designs a method to realize the reading and information extraction and the generation of description text of the content represented by regional classified statistic graph method based on the color determination results and the elements detection results of the map title and the legend.Experiments prove that the method is effective.The correct rate of administrative division unit matching level reaches 99.09%.(4)This paper proposes a multi-scale model fusion strategy to detect the constituent symbols of the statistical chart of the "change" method.On the scale of the statistical chart of the "change" method,the data set of constituent symbols is produced and the Faster R-CNN model is trained to detect.The m AP of the model on the validation set is 90.91%.The object detection models of the two scales of the regional statistical chart and the constituent symbols are combined in series to realize the detection of constituent symbols on the scale of thematic map.The comparison with single-scale model detection and template matching methods proves the effectiveness and superiority of the multi-scale model fusion strategy to detect the constituent symbols of the statistical chart.(5)This paper proposes a method to determine the administrative division unit of the statistical chart according to the boundary prediction boxes of the statistical chart and the administrative division unit.The accuracy rate of the experiment on the data set reaches 100%.This paper designs a method to realize the reading and information extraction and the generation of description text of the content represented by cartodiagram method based on the detection result of the composition symbol of the statistical chart,the determination result of the administrative division unit of the chart and the elements detection results of the map title and the legend.The experimental results show a high degree of accuracy.