Comparative evaluation of machine learning models for museum exhibit recognition from video-derived datasets

This study evaluates the performance of multiple deep learning models for automatic recognition of museum artifacts using image frames extracted from real-world video footage. A comparative analysis is conducted across eight state-of-the-art architectures—MobileNetV3, ResNetV2, EfficientNetV2, You Only Look Once v8 (YOLOv8), Visual Geometry Group 16 (VGG16), ConvNeXtTiny, SwinV2-Base, and Dual Attention Vision Transformer (DaViT)—on a custom dataset collected in an actual museum environment. The dataset comprises labeled video frames categorized by artifact type and is used to train and test models for both classification and object detection tasks. Results indicate that YOLOv8, MobileNetV3, and DaViT achieve superior performance for real-time mobile and augmented reality (AR) applications, while ResNetV2 and SwinV2-Base provide high classification accuracy suitable for archival and cataloging systems. This work offers practical guidance on dataset design, model choice, and deployment strategies for artificial intelligence (AI)-powered cultural heritage technologies.