Novel algorithm for video summarization using deep learning rnns

Abstract

Due to the huge development of automated video processing newlinetechniques, the video signal has seen exponential growth. If the end-user is newlineinvolved only in some important part, it is a memory wastage and time newlinewastage for video databases to store a complete video. Time consumption can newlinebe reduced remarkably by watching the summarized video before the actual newlinevideo. To reduce this limitation, we use video summarization. In Video newlineSummarization, an abstract view of entire video is a processed within a short newlineperiod of time. The summaries of the videos are generated in this technique. newlineThose summaries contain maximum information to make the user to newlineunderstand more easily. In this work, one of the deep learning neural newlinenetworks, Recurrent Neural Network (RNN) and its variants are used to form newlinevideo summaries. The performance results of the new methods using Multi- newlineEdge optimized Long Sort Term Memory (LSTM) RNN, Hierarchical newlineMultiscale LSTM (HM-LSTM) RNN and Error Correction RNN (EC-RNN) newlineare implemented. Each technique is tested to provide the video summary for newlinethe standard datasets such as MED data set, VSUMM data set and YouTube newlinedata set. All the proposed algorithms are evaluated in terms of Precision, newlineRecall, F-Score and Average Processing time and the proposed algorithms newlineproduced better performance than the existing methods. The results show the newlineimprovement in higher percentage in Precision, Recall, F-Score and lower newlinevalue in Average Processing time from algorithm to algorithm. It is further newlineobserved that the video summarization resulted from the proposed algorithms newlineare having wide scope in video surveillance, video retrieval and video newlineindexing applications. newline