Automating Cultural Memory: AI Image Captioning for The National Museum of Qatar's Photographic Collection

Abstract

Cultural heritage institutions face a growing metadata bottleneck: digitization has scaled rapidly, but expert description has not. This thesis investigates AI-based image captioning for the photographic collection of the National Museum of Qatar (NMoQ), addressing three research questions: the impact of prompt configuration on caption quality (RQ1), the influence of model training background on cultural awareness (RQ2), and the validity of LLM-as-judge evaluation in this cultural context (RQ3). Five prompts were tested on a 10-image pilot using GPT-4o mini, and the best-performing prompt was applied across three models: GPT-4o mini, Fanar Oryx-IVU-2, and Qwen 2.5-7B, generating 1,350 captions from a 450-image dataset. A 150-caption reference subset was manually evaluated, and the full dataset was scored using Gemini 2.5 Pro as an automated judge calibrated against expert scores. Structured prompt engineering raised Cultural Awareness scores from 2.70 to 4.10 without fine-tuning. GPT-4o mini outperformed both regional models across all evaluation methods. Fanar's Gulf-specific training produced no statistically significant advantage in cultural awareness over Qwen 2.5 under expert evaluation (p = 1.000), challenging the assumption that regional training translates into improved cultural captioning in English-language tasks. The Geminii judge achieved moderate post-refinement agreement (Pearson r = 0.481) but exhibited a systematic leniency bias toward Fanar that rubric instructions alone could not resolve. The study contributes an evidence-based prompting framework for Gulf heritage captioning, empirical evidence against the assumed value of regional model training for English-language cultural tasks, and a replicable LLM-as-judge calibration methodology for non-Western heritage contexts.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Humanities and Social Science