Video Frame Prediction Using Convolutional LSTM Encoder and Decoder

With rapid growth of audiovisual content now spreading through social networking sites, chances of exposure towards violent and mature content by the youth are growing. This project addresses the risk of detection of such content using advanced video frame prediction techniques. We present here an encoder-decoder model of Convolutional LSTMs in the film for violent and obscene sequence detection. The spatiotemporal nature of video data encourages our algorithm to exploit accurate predictions of future frames from past content, which can be used in real time for dangerous material detection. The study relies on recent literature that highlights the psychological implications arising from the viewing of violent content by teenagers, thus making the need for efficient automated detection mechanisms ever more pertinent. Furthermore, we present a complete definition of “violence” for the standardized detection mechanisms and to cross-compare various studies. Using Violent Scene Detection, our results indicate high improvements in the detection accuracy and therefore become effective in this context for parental control and content moderation. Such results help safeguard children in the increasingly digital environment: traditional as well as currently ever-increasing.