Visual Feature Extraction and Machine Learning for Graphical Violence Detection: A Deep Learning-Free, Efficient Approach

This study proposes an efficient, deep learning-free approach for detecting graphically violent images to address the challenges of high computational costs and class imbalance in digital content moderation. Using the "Graphical Violence and Safe Images" dataset, we employed a hybrid feature extraction strategy combining color (3D Histogram), texture (Local Binary Patterns [LBP], Gray-Level Co-occurrence Matrix [GLCM]), and shape (Histogram of Oriented Gradients [HOG]) descriptors, followed by Analysis of Variance (ANOVA)-based feature selection. Among five machine learning models evaluated, XGBoost achieved the highest performance with 96.55% accuracy and an 84.38% Macro F1-Score on the test set. Furthermore, the proposed method offers a processing time of approximately 33.85 ms per image on a standard CPU. The results demonstrate that the proposed method offers a computationally efficient and interpretable alternative to deep learning for real-time applications.