A hybrid approach based on visual transformer-based feature extraction and machine learning classifiers for object recognition in unbalanced image datasets

This study aims to improve object recognition performance by combining features obtained from the Vision Transformer (ViT) model with classical machine learning classifiers such as LightGBM, AdaBoost, ExtraTrees, and Logistic Regression. In the research, different feature extractor-classifier combinations were compared as a result of experiments conducted on the Caltech-101 dataset. In particular, the combination of ViT and Logistic Regression stood out as the recommended method, achieving 95.5% accuracy and 89.7% sensitivity. These results demonstrate that the proposed approach exhibits significantly higher performance compared to existing cutting-edge methods. The findings show that synthesizing features extracted from the ViT model with classical classifiers can lead to significant improvements in object recognition tasks. The proposed hybrid method offers high representational power for deep learning-based automated feature extraction; By combining the advantages of traditional machine learning classifiers, such as intelligibility and low computational cost, it has the potential to deliver significant improvements in accuracy, processing performance, and overall system efficiency. Furthermore, it stands out as an effective alternative solution for real-time analytics and intelligent system applications with limited hardware.