An artificial intelligence method for text placement evaluation in maps

إنجليزي. A map is a combinatorial presentation of map objects and labels providing users with sufficient and useful information. Text placement is the most time-consuming and cumbersome work in a map production process, and many methods have been proposed to automate this process. However, text placement evaluation lacks exploration, especially when artificial intelligence (AI) methods are experiencing prosperous development. To bridge the gap, this master thesis proposed an AI method for a text placement evaluation. Firstly, London street maps were simplified and rasterized based on cartographic rules formulated by experienced cartographers. Secondly, a text placement quality measurement schema was designed. Thirdly, datasets were prepared by manually classified rasterized maps into different quality levels. Fourthly, GoogLeNet, a widely accepted convolutional neural network, was trained, and its performance was validated. The results showed the GoogLeNet model’s poor training accuracy and validation accuracy in the training process and bad performance in evaluating text placement. All the test images were classified as having bad text placement. These pieces of evidence indicated that the trained GoogLeNet model was not reliable and qualified for evaluating text placement at this stage. The negative results might be caused by the capacity issue, dataset issues and model issues. These issues were discussed in the latter part of the report. Finally, some limitations and perspectives of using AI for text evaluation were listed in the end for further enlightenment.

إنجليزي. A city map is the epitome of urban space, which aims to provide citizens and tourists information about the urban environment. It has become an inseparable part of daily life in many fields, such as transportation, tourism and navigation. Some critical objects on the map are labelled with texts or icons to indicate their locations and identities. Such efficient and clear presentations of crucial information on the map make a user obtain the surrounding information quickly and thus make the city map valuable. Therefore, high-quality city maps are indispensable in the urban public environment. However, it is impossible to achieve a high-quality city map without good text placement. Placing texts at good positions to indicate their referent is the most time-consuming and cumbersome work in map production. Many approaches have been proposed to automate this process for efficiency. However, less attention was paid on text placement evaluation, especially when artificial intelligence (AI) methods are experiencing prosperous development. To bridge this gap, this master thesis proposes an AI method for a text placement evaluation. Firstly, some map objects are selected and retained based on cartographic rules formulated by experienced cartographers. Symbologies of their text labels are simplified, and maps are rasterized as images. The second step is designing a measurement schema for evaluating text placement quality level, which is followed by a manual classification of rasterized images based on the schema. These images are given various quality levels and split into different datasets. Finally, GoogLeNet, a widely accepted convolutional neural network, is trained, verified and assessed with prepared datasets. The results show the model’s poor accuracies in the training process and bad performance in evaluating text placement. Test images, regardless of manual quality level classification results, are all classified as having bad text placement. These pieces of evidence indicate that the trained GoogLeNet model is not reliable and qualified for evaluating text placement at this stage. The negative results may be caused by the capacity issue, dataset issues and model issues. These issues are discussed in the latter part of the thesis. Finally, some limitations and perspectives of using AI for text placement evaluation are listed in the end for further enlightenment.