Computational Comparison of CNN Based Methods for Violence Detection

Abstract

Abstract In this paper, we approach with four different CNN-based models i.e., VGG-19, VGG-16, InceptionV3 and MobileNetV3 with an improved version of the previous models for violence detection and recognition from videos. The proposed models use the pre-trained models as the base model for feature extraction and for classification after freezing the rest of the layer, the head model is prepared with averagepooling2D of (5, 5), and after flattening only one dense layer having 512 nodes with ‘ReLU’ activation function, dropout layer of 0.5 and last output layer with only 2 classes and ‘softmax’ activation function. This head model of fully connected layers was used in the proposed models. These models are trained and evaluated on the Hockey fight dataset and Real life violence situations detection datasets. The experimental results are far better in terms of accuracy and other performance metrics and the models have reduced parameters and less computational time than previous models.