Tarsius Image-Source Classification Using VGG Feature Extraction and Machine Learning Classifiers
TL;DR
VGG-19 with Logistic Regression provides the most balanced performance for the available dataset and demonstrates that predictive metrics and feature-space visualizations should be interpreted jointly and remain limited by the small, web-sourced dataset and incomplete provenance records.
Abstract
Background: Reliable classification of heterogeneous Tarsius imagery is relevant to digital wildlife-data management, yet real photographs and AI-generated images may share visual characteristics that complicate automated separation. Objective: This study compares VGG-16, VGG-19, and InceptionV3 feature extraction combined with Logistic Regression and Neural Network classifiers for real, cartoon-style, and AI-generated Tarsius images. Methods: A balanced dataset of 300 web-sourced images (100 per category) was processed in Orange Data Mining. Performance was evaluated using AUC, classification accuracy (CA), F1-score, precision, recall, Matthews Correlation Coefficient (MCC), confusion matrices, Multidimensional Scaling (MDS), and silhouette plots. Results: Logistic Regression with VGG-19 produced the strongest reported combination (AUC = 0.997, CA = 0.967, F1 = 0.967, precision = 0.968, recall = 0.967, and MCC = 0.950), closely followed by VGG-16 with Logistic Regression. Cartoon-style images were generally the most distinguishable, whereas real and AI-generated images showed the greatest overlap. InceptionV3 produced clearer two-dimensional MDS separation but lower aggregate predictive performance than the VGG-based Logistic Regression configurations. Conclusion: VGG-19 with Logistic Regression provides the most balanced performance for the available dataset. The findings demonstrate that predictive metrics and feature-space visualizations should be interpreted jointly and remain limited by the small, web-sourced dataset and incomplete provenance records.