Elly Matul Imah, Karisma
Violence against children is a severe problem. Violence causes physical and mental trauma and can even threaten the lives of victims, especially children. Therefore, violent cases need special attention and require detection in their handling. Violence detection research is still a challenge for researchers and a considerable effort. A training process on video datasets is extensive empirical studies. Finding the optimal feature set and classifier is needed to achieve good recognition results. This paper presents violence detection using the visual geometry group network-16 (VGGNet-16)-based deep transfer learning feature extraction, combined with ensemble decision fusion learning. Ensemble decision fusion learning is a kind of ensemble learning method. It combines classifiers from multiple models and datasets. The majority of voting connects the classifier's output is used to decision fusion in this study. The majority voting counts the votes of the base learners and predicts the final class as an output; it is less biased toward the outcome. The deep learning classification methods used as an ensemble are LSTM, BiLSTM, GRU, and SVM. The experimental results show that a combination of VGGNet-16 and ensemble decision fusion learning can increase children's violence detection accuracy on surveillance videos. The obtained accuracy is 92.4% better 1.5% to 22.7% among other methods. © 2022. All Rights Reserved.
Department of Mathematics, Faculty of Mathematics and Natural Sciences, Universitas Negeri Surabaya
