A large-scale near-duplicate video retrieval method based on random multi-view hashing

Abstract

The invention discloses a large-scale repeated video retrieval method based on random multi-view hash. Video key frames are extracted through a time interval sampling method, and color histogram HSV and local binary pattern LBP characteristics of the key frames are extracted; the HSV and LBP characteristics are mapped to quasi-hash space through linear mapping and the sigmoid function, and hash codes representing videos finally are obtained through quasi-hash codes generated through thresholding processing; the Hamming distance between a test video hash code sequence and other video hash code sequences is calculated, and approximate repeated videos of the test video are retrieved according to the Hamming distance. Multi-view characteristic information of the videos is used, retrieval efficiency and accuracy are improved through hash mapping, and large-scale quick retrieval of the approximate repeated videos is achieved.

Description

technical field [0001] The invention belongs to the fields of video retrieval and machine learning, and specifically relates to a large-scale near-repetitive video retrieval method based on random multi-view hashing. Background technique [0002] In the 21st century when the Internet is highly developed, a large number of video-related applications and services continue to emerge on the Internet, such as video sharing, video recommendation, and video broadcasting. The Internet is full of massive video data and shows a trend of rapid growth. Rich video content attracts more and more network users to edit, upload, download and search various videos. According to comScore, in August 2014 alone, computer users browsed nearly 300 billion videos, with an average of 202 videos per user, totaling 952 minutes. Studies have shown that there are a large number of potential near-duplicate videos in such a huge amount of videos. Therefore, how to retrieve and remove these near-duplicat...

AI Technical Summary

Problems solved by technology

At present, some methods that completely compare the differences between all key frames of each pair of videos, and then calculate the ratio of similar key frame pairs to determine whether it is a near-duplicate video have achieved good retrieval results, but the retrieval speed is far behind. Cannot meet engineering requirements

And some algorithms that generate video content tags are faster, but the effect still needs to be improved

There are also some algorithms that use retrieval structures such as trees and hashes. Although they have overcome the defects of the above two methods, they still have limitations in retrieval accuracy and scalability, and most of them can only use a single key frame. Feature information, such as global features or local features

Images