En jämförande studie av GPT-4o och ett modernt ALPR-systems prestanda på multinationella registreringsskyltar : Utvärdering av modellens noggrannhet för olika landspecifika registreringsskyltformat

In an increasingly globalized context, where traffic is becoming more populated by vehicles from different parts of the world, ALPR systems often struggle to accurately identify unfamiliar country-specific license plate formats and designs. This study evaluates the accuracy performance of the multimodal AI model GPT-4o, with its advanced vision capabilities, against a state-of-the-art ALPR system, for the challenging task of identifying diverse multinational license plates. Both models were tested on a diverse range of datasets with images containing plates from nine countries, including European, Eurasian, and South American countries. For evaluation, the models were tested on plate-level; measuring accuracy as the percentage of correctly predicted license plates out of the total ground truth, and on character-level measuring the error rate on individual license plate characters. A Mcnemar test, z-test, and paired t-tests was applied to assess the statistical significance of the performance differences. Results show that GPT-4o achieves overall accuracy comparable to the ALPR, and in some cases superior, especially when recognizing country-specific license plate formats unfamiliar to the ALPR. This highlights the potential of GPT-4o in solving tasks in the license plate recognition domain, without the necessity of fine-tuning or training on specific formats. However, previous studies show that some fine-tuned real-time ALPR solutions will achieve superior performance, making GPT-4o a promising solution as a fallback on the employed ALPR, in environments where license plate formats can be expected to be diverse.