Recurrent unit augmented memory network for video summarisation

Video summarisation can relieve the pressure on video storage, transmission, archiving, and retrieval caused by explosive growth of online videos in recent years. Most existing supervised methods use convolutional neural network (CNN) or recurrent (RNN) to model temporal dependencies between frames shots. CNN mainly focuses local information, RNN loses long-term information when input sequence is long, both which have limited ability obtain long-range memory video. Therefore, a unit augmented (RUAMN) for proposed, effectively utilises extraction end-to-end (MemN2N) solves problem that MemN2N insensitive information. At same time, proposed RUAMN enhances process update multiple computational steps (hops), finally generates meaningful result. Specifically, composed module, global-and-local sampling, module output module. The uses bidirectional GRU forward backward each frame. Then sampling performs global respectively several shorter sequences, so modules capture fine-grained relationship features more effectively. extracts feature sequence, frame-level importance scores are predicted Extensive experiments benchmark datasets, is, TVSum SumMe, demonstrate superiority our method over state-of-the-art methods.