Empowering a Violence Prevention with CNN-SVM based Classification

This research study comprehensively investigates the violence categorization using a hybrid Convolutional Neural Network-Support Vector Machine (CNN-SVM) model for five violence classes. We aim to create a model that can precisely and correctly discriminate between physical, verbal, psychological, sexual, and property violence. The model achieved class-specific accuracy measures from 89.38% to 95.28%, recall rates from 85.98% to 97.66%, and F1 scores from 90.02% to 95.35% as per the experimental results. A high accuracy of 92.6559% indicates the model's resilience. We employed multiple averaging methods to evaluate the general performance of the model. The macro averages of the accuracy, recall, and F1-score were 92.65%, 92.85%, and 92.69%, respectively, indicating similar performance across classes independent of classimbalance. The accuracy, recall, and F1-score of 92.74%, 92.66%, and 92.63%, respectively, were calculated considering the support of each class. This method provides a more comprehensive assessment of the model's performance since it thinks the number of samples in each class. Accuracy, recall, and F1-score micro averages were always at 92.66%, which indicates the model's ability to classify violence well, irrespective of class distribution. The measurements show that the CNN-SVM method effectively classifies complex, multi-class violent images. This study improves the research area for automated content moderation and monitoring by offering a detailed analysis of recognizing violent content in various settings and media.