An Empirical Study of CNN-LSTM on Class Imbalance Datasets for Violence Video Detection
Moch Arief Soeleman, Catur Supriyanto, Dwi Puji Prabowo · 2021
Violence detection has become an important topic of video surveillance in the last decade. Some studies in violence video detection demonstrated that learned features from Convolution Neural Network (CNN) gives high accuracy compared to handcrafted features. For this reason, we evaluate several CNN architectures to detect violence action in video. This work compares five pretrained networks VGG16, VGG19, ResNet50, Inception V3, and Xception. Then, the extracted features from each frame are forwarded to a long short-term memory (LSTM) network. We evaluate the pretrained networks on class imbalance datasets since violence video detection might suffer from class imbalance. Two public datasets are being used to evaluate the model; hockey fight dataset and violent crowd dataset. Our experiment results show that InceptionV3 achieved better performance in most cases.