Využití Multimodálních Velkých Jazykových Modelu pro analýzu fotografií

This thesis investigates the potential of multimodal large language models to be used as a tool for the official automated passport photo classification, an area where high accuracy is demanded. The publicly available MMLLM (Multi-Modal Large Language Model) LLaVa v1.6 Vicuna 13B model was used and adapted using different tactics. The primary objective was to create a model that could classify passport-style photographs according to criteria defined by an international standard. A manually curated dataset was created to serve both as a proof-of-concept and as training material for fine-tuning. In the process of fine-tuning, the learning rate and the number of epochs were systematically varied to achieve the best possible result. The fine-tuned model achieved both F1 and AUC of 0.77. These results demonstrate measurable improvements over the base model. Further improvements are suggested through enhanced dataset diversity and more granular supervision.