Scholarlink

Accurate georeferencing of scanned topographic maps is vital for modern Geographic Information Systems (GIS). We introduce an end-to-end pipeline that combines an Optical Character Recognition model (EasyOCR) with a two-stage Multimodal Large Language Model (LLM) to extract and parse printed latitude/longitude labels. High-resolution PDFs (Survey of India, 1:50 000) are converted to 700 DPI PNGs and cropped to margins. EasyOCR provides text and bounding-box data; the first LLM prompt filters and merges detections, and the second enforces consistency, outputting degree–minute–second pairs. Post-processing fits an affine transform with RMSE-based outlier removal to generate ground control points. GDAL then creates GeoTIFFs in EPSG:4326. Tested on twelve sheets, the OCR+LLM workflow yields internal RMSE <5 m and validation RMSE <30 m on eleven maps, reducing errors by 2–4 orders of magnitude compared to an LLM-only baseline while processing each sheet in under 350 s. This open-source solution scales cartographic digitization and unlocks geospatial archives. Code link is also shared. ( https://github.com/pallavi093/OCR_LLM_Georeferencing

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated Georeferencing of Topographic Maps via OCR and In-Context Multimodal LLM Reasoning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pallavi Tyagi,
  • Vishal Dubey

摘要

Accurate georeferencing of scanned topographic maps is vital for modern Geographic Information Systems (GIS). We introduce an end-to-end pipeline that combines an Optical Character Recognition model (EasyOCR) with a two-stage Multimodal Large Language Model (LLM) to extract and parse printed latitude/longitude labels. High-resolution PDFs (Survey of India, 1:50 000) are converted to 700 DPI PNGs and cropped to margins. EasyOCR provides text and bounding-box data; the first LLM prompt filters and merges detections, and the second enforces consistency, outputting degree–minute–second pairs. Post-processing fits an affine transform with RMSE-based outlier removal to generate ground control points. GDAL then creates GeoTIFFs in EPSG:4326. Tested on twelve sheets, the OCR+LLM workflow yields internal RMSE <5 m and validation RMSE <30 m on eleven maps, reducing errors by 2–4 orders of magnitude compared to an LLM-only baseline while processing each sheet in under 350 s. This open-source solution scales cartographic digitization and unlocks geospatial archives. Code link is also shared. ( https://github.com/pallavi093/OCR_LLM_Georeferencing