Reading index cards
Structured fields are generated from scanned pages of cards.
Data Generation from the Historical Analog Inventory of the Zentralinstitut
The historical analog inventory of the Photothek, which spans from the institute’s founding years to 1999, consists in its early part of five handwritten ledger volumes as well as 66,000 pages containing approx. 529,000 cards, documenting around 700,000 inventory numbers.
The goal of the processing is to automatically read each scanned page of cards and break down its contents into machine-readable fields — artist, location, work description, catalog information, and photo credit.
Instead of classical text recognition (OCR), which often fails on handwritten notes and faded ink, a multimodal AI helps to semantically identify and separate the contents. This approach has proven useful in the project because it also recognizes content that is not placed according to a standard layout and can handle the index card layouts that vary slightly over the decades.
Technical Approach
The input consists of JPG scans, each showing eight cards in a 2×4 grid. Each image is sent to Google Gemini 2.5 Flash; the prompt instructs the model to return five fields per card as JSON: artist/school, location, work description, catalog info, photo credit.
The central principle is content-first prompting: the model is instructed to first classify content-wise (“This is an artist’s name”), and to use the position on the card only as secondary validation. This keeps extraction stable even when layouts vary or fields are slightly misaligned.
Processing runs in five-way parallelism with adaptive throttling to comply with the Gemini API quotas. Results are saved per image as JSON, then exported into a combined file per series as well as to CSV and Google Sheets.
Data Sources & Tools
- Input: JPG scans of the ZI Photothek (2×4 grid)
- Model: Google Gemini 2.5 Flash (multimodal)
- Output: JSON per card, combined JSON per series, CSV, Google Sheets
- Cost: approx. €0.002–0.005 per scan
- Repository:
inventories_prod(internal)