Recognizing texts on maps

Current popular optical character recognition(OCR) models struggle to achieve accurate results in both text positioning and text labelling prediction due to the complex, diverse and noisy nature of historical maps. Oftentimes, text-bounding polygons in historical maps intersect each other, and te...

Full description

Saved in:
Bibliographic Details
Main Author: Tan, Pheng Khai
Other Authors: Li Boyang
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2024
Subjects:
Online Access:https://hdl.handle.net/10356/175181
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Current popular optical character recognition(OCR) models struggle to achieve accurate results in both text positioning and text labelling prediction due to the complex, diverse and noisy nature of historical maps. Oftentimes, text-bounding polygons in historical maps intersect each other, and texts are overlaid on top of surrounding geographical features, which add unnecessary noise and difficulties to the task of recognizing texts from maps. This paper presents a method to automatically generate large amounts of annotated training data by using CycleGAN to generate synthetic historical maps. The generated data is then used to train a DPText-DETR model, a model selected due to its distinct feature that gives it the potential to excel at the task required for this project. A pipeline is then proposed to be implemented to make historical map OCR more accessible and user-friendly. In this paper, thorough analysis and evaluation have been conducted on the proposed method, comparing it against a baseline model that adequately represents the pros and cons present in modern OCR tools, as well as a state-of-the-art model. Our evaluations show that this approach not only simplifies the generation of annotated training data but also significantly enhances OCR accuracy. Comparative performance assessments reveal that our model achieves a precision increase of 4.25 %, recall by 2.58%, and overall F1 improvement by 3.23% over baseline and state-of-the-art models in terms of Wolff’s metric, setting a new benchmark for historical map text recognition.