Presentation Information
[A-13-06]Evaluation of Vision-Language Models for Vehicle Model and License Plate Recognition
◎Koki Niwa1, Chiyomi Miyajima1 (1. Daido University)
Keywords:
Vehicle model recognition,License plate recognition,Vision-language model
Accurate vehicle identification is essential for traffic surveillance and public security. While license plate information is widely used, combining it with vehicle model information can improve identification reliability. This study evaluates seven Vision-Language Models (VLMs) for vehicle model and license plate recognition using Japanese vehicle images. Vehicle and license plate regions were extracted by object detection, and the vehicle make, model, body color, and license plate information were recognized. Experiments were conducted using 161 front-view vehicle images. The results showed that the cloud-based VLMs and some local VLMs achieved over 90% accuracy in vehicle make and body color recognition, whereas Vehicle model recognition proved to be more challenging, with the best local model achieving 43%. In contrast, license plate recognition exceeded 99% for the cloud-based VLMs and 98% for a local VLM. Furthermore, some VLMs, such as Qwen3.5, showed less performance degradation in low-resolution scenarios when fed the cropped license plate region.