Arabic geospatial text spotting improves with tool-augmented language models
ArabGeoText: Tool-Augmented Vision-Language Models for Arabic Geospatial Text Spotting
Computer Vision and Pattern Recognition
Summary
Finding and reading Arabic place names on maps is hard because the words are small, close together, and turned in different directions. The authors created a new dataset of Arabic geographic labels from two cities, pairing map renderings with satellite images to help computers learn. They tested a new approach that uses a tool to find text areas first, then a language model fine-tuned to read the text more accurately. Their method works better than just trying to read everything at once, making it easier to spot Arabic place names on complex map images.
What this means in practice
- •For mapping and gis teams: Improve automatic extraction of Arabic place names from maps and satellite images to update geographic databases with accurate text labels.
- •For urban planners and local governments: Use better geospatial text recognition to support city planning and navigation systems that rely on accurate Arabic map labels.
Authors
Yakoub Bazi, Sarah Alsaleh, Mohamad M. Al Rahhal, Mohamed A. Mekhtiche, Mansour Zuair
Abstract
Geospatial text spotting involves locating and transcribing geographic labels, such as road names, district names, and landmarks, in raster maps. These labels are often small, densely arranged, and arbitrarily oriented. Although Arabic text recognition has been studied extensively in documents and natural scenes, Arabic geospatial text spotting remains underexplored. We introduce ARABGEOTEXT, an Arabic-centric benchmark covering 970 geographic locations in Riyadh and Jeddah, with 17,875 manually annotated text instances per view type. For each location, we pair an OpenStreetMap (OSM) rendering with a composite created by overlaying that rendering on the corresponding remote sensing (RS) imagery. The paired views preserve label content and pixel geometry, allowing evaluation across different backgrounds. Geographic partitions support in-city and city-disjoint evaluation of detection, recognition, and end-to-end spotting. Experiments with direct Qwen3.5-4B spotting show difficulties in recovering all labels within a tile. We therefore propose a tool-augmented vision-language framework that combines YOLO26 localization and crop rectification with Qwen3.5-4B recognition adapted through low-rank adaptation (LoRA). The framework outperforms direct LoRA-adapted Qwen3.5-4B spotting across both image views and test cities. With localization held fixed, reader adaptation further improves recognition over the pretrained zero-shot reader. The dataset and code will be made publicly available at https://github.com/yakoubbazi/ArabGeoText.