Automatic Tag Generation (ATG) Using Machine Learning Tech-niques for Women Violence Detection

Abstract

This research addresses the challenge of automated detection and classification of multiple types of violence from image data using deep learning techniques. Given the societal importance of timely and accurate violence recognition, this study explores both custom Convolutional Neural Network (CNN) architectures and state-of-the-art transfer learning models pre-trained on ImageNet, including VGG16 and InceptionV3. The dataset comprises images across five violence categories—brutality, domestic violence, human trafficking, rape, and sexual harassment—collected and augmented to enhance model generalizability. Methods involved image preprocessing, data augmentation, and training models with categorical cross-entropy loss optimized via Adam. Transfer learning approaches outperformed the custom CNN, The models demonstrated varying degrees of success in classifying violence image categories. Transfer learning models, particularly VGG16 and InceptionV3, outperformed the custom CNN, achieving overall accuracy improvements from approximately 75% to 76%. These results confirm the effectiveness of leveraging pre-trained networks for complex image classification tasks with limited datasets.

Class-wise analysis through confusion matrices and derived metrics such as precision, recall, and F1-score demonstrated varied detection performance, highlighting difficulties in differentiating visually similar classes. The results affirm that leveraging pre-trained deep architectures substantially benefits the classification of limited, complex image datasets. This paper contributes by providing a comprehensive evaluation of deep learning approaches for violence classification in images, motivating their use in practical monitoring and intervention applications. Future work is suggested to integrate temporal data and attention mechanisms to further enhance detection performance. The findings underscore the feasibility and importance of automated violence recognition systems for social safety.