Comparative Analysis of Deep CNN and VGG-16 for Anomaly Detection

In video surveillance systems, anomaly detection is a crucial task for which deep learning models present potential answers. This study examines how well a custom deep Convolutional Neural Network (CNN) performs for anomaly identification in surveillance flicks in comparison to the widely used VGG-16 architecture. There are fourteen classes in the dataset, encompassing both typical activity and unusual oddities. Using common metrics like accuracy and loss, both models are trained and assessed. The custom CNN performs better than the VGG-16 (99.58 %) in terms of accuracy (99.92 %), suggesting that it is more successful in obtaining pertinent features for anomaly identification. Even while VGG-16 is well-known for its ability to classify images, in this particular anomaly detection scenario, its performance is inferior to that of the custom CNN. This conclusion emphasizes how crucial it is to customize model architectures for anomaly detection applications to match the specifics of the dataset and the task at hand in order to get the best possible performance.