SecureVision: A CNN-Based Review of Deep Learning Frameworks for Women's Safety

Abstract

This paper presents SecureVision, an evaluation of the use of deep learning to enhance women&s;s safety by detecting visible threats. This will include an in-depth discussion of Convolutional Neural Networks (CNNs) and identifying violent and aggressive behavior based on surveillance videos. The evaluation will explain different such CNN architectures including MobileNetV2, InceptionResNetV2 and ResNet, and share test results based on testing datasets like the Annotated Real-Life Violence Situations Dataset. A number of commonly used preprocessing approaches are listed with some detail, like: image augmentation; image resizing, and normalization, as steps to improve model robustness. Unlike deployable safety apps, SecureVision is devoted to academic evaluation of model performance metrics like: accuracy; precision; recall, and the F1-score. Security also includes further descriptions of the weaknesses of uni-modal detection series from earlier discussion, and an excursion into future development of multimodal systems. The research is useful to baseline studies of AI-enhanced safety apps, and provides pathways to supporting builds for more autonomic, real-time threat detection.