Researches On Efficient Hashing For Video Retrieval

In recent years, with the development of multimedia and Internet technology, videos have increasingly become the highlight of information media due to their rich content. Until December 2015, the number of Internet video users of China has reached 504 million and thanks to the high computing ability of mobile terminal and improvement of network bandwidth, the number of mobile phone users has exceeded 405 million. According to statistics, about 7.5PB data is produced each minute on the Internet all over the world, of which about 90 percent are videos and images and it continues to increase. In the face of such large-scale data, with the limited bandwidth and computational cost, it has important impact on the field of social information to obtain the required information quickly and accurately via efficient video retrieval. As one of the core technologies of video retrieval, video hash technology has been concerned by a great many scholars at home and abroad.This paper starts with exploring the memory features of images and then proves that features based on visual attention have positive influence on predicting image memorability. Then the memory features are adopted to represent the video content and a memorability feature based video hashing is proposed as an alternative to appearance feature and visual attention based algorithms. At last, given the fact that determining hash length in the current video hashing algorithms has been neglected, a method of hash length prediction for video hashing is proposed.The main innovation and contribution of this paper are showed in the following three aspects:(1) The influence of visual attention on image memorability is investigated. Exploring the effective features to predict image memorability is one of the core issues of the researches on image memorability. From our perspective, appearance features are on the first step and it is they that stimulate the vision. While visual attention modelling simulates the behavior of human visual system by automatically producing saliency map of the target image. The experiments are carried out by using two existing visual attention models and demonstrate that these mentioned visual attention based features are more effective than the appearance features to predict the image memorability.(2) A memorability based method for video hashing is proposed. A memorability feature based video hashing is proposed as an alternative to appearance feature and visual attention based algorithms. Experiments on different kinds of videos demonstrate that the proposed MF-Hashing algorithm is promising in video hashing.(3) Hash length prediction for video hashing is proposed. In this proposed method, the efficient hash length of video hash can be determined simply, in which the approximate optimal hash length of the whole dataset can be determined only through the training dataset. This can provide a reference for the hash length in the whole dataset. Experiments on two video datasets prove that the optimal hash length predicted by the proposed method is feasible and reliable.This paper not only investigates the influence of visual attention on image memorability but also apply this result to video hashing algorithm and a memorability feature based video hashing is proposed. The researches of video hashing are from the view of extracting features, which include appearance features that stimulate the vision and combine those from the perception and memory layers. In addition, hash length prediction for video hashing is proposed, which provides a reference for the hash length in the whole dataset. In conclusion, the effective representation of video content, not only is advantageous to the efficient video retrieval in theory, but also can be applied in the field of public safety, video sites, mobile search, etc.