Developing BrutNet: A New Deep CNN Model with GRU for Realtime Violence Detection

Computer vision with deep learning has recently emerged for Automatic Violence Detection and Classification (AVDC) with enormous potential. This paper reports an early development of a new Deep Convolutional Neural Network (DCNN) model that we call BrutNet. Building on the Gated Recurrent Unit (GRU), the BrutNet is designed to operate on the patterns within multiple frames of a video or video clips of shape 160× 90 with a duration of at least 3 seconds. For obtaining the image-feature set and the pattern of each frame, convolutional layers were considered for each frame of the time distributed layer. The model thus encodes the data from 4D to 2D to obtain a 512-features set for each frame. The temporal nature of these frames is then extracted by the GRU layer as a 1D vector, which is processed by several dense layers. A binary classification is thereby performed denoting the content as violent and non-violent. Dropout layers with a dropping rate of 0.25 were added to avoid overfitting the model. Besides, ReLu-activation and sigmoid-activation functions were defined in the hidden and output layers, respectively. Being trained with a recent high-resolution AVDC video dataset and appropriate hyper-parameters on the NVIDIA Tesla K80 GPU of Google Colab, the initial testing and validation of the model has recorded a test accuracy of 90.00% outperforming the earlier LSTM based ResNet50 model.