Unsupervised Reinforcement Learning For Video Summarization Reward Function

We propose a new reward function based on Deep Summarization Network (DSN), which is used to synthesize short video summaries to facilitate large-scale browsing of videos. The DSN uses the video summarization as a process of sequential decision making, predicting the probability of each video frame to indicate the likelihood that the video frame is selected, and then selecting the frame based on the probability distribution to form video summaries. By designing a new DSN reward function, the rewards for representative and diversity rewards are higher, and a large number of experiments are performed on the two benchmark datasets, demonstrating that our summary network is significantly better than existing unsupervised video summaries.