Near-duplicate Video Retrieval Based on Deep Unsupervised Key Frame Hashing

Abstract

In recent years, original videos are often re-edited, modified and redistributed, which not only cause copyright problems, but also deteriorate users’ experience. Near-duplicate video retrieval based on learning to hash has been widely concerned by people. However, there are still two major defects with existing methods. Firstly, the information capacity of hash code needs to be maximized. Secondly, the retrieval efficiency of partially repeated video is insufficient. In this paper, we propose a near-duplicate video retrieval method based on key frame hashing to improve retrieval performance. We design the semi-distributed hash layer to force the distribution of the continuous key frame hash code to approach the optimal distribution, i.e., the half-half distribution. By minimizing the semantic loss, quantization loss, and bit uncorrelated loss, we train our model to generate compact binary hash codes. To retrieve partially repeated videos, the proposed video subsequence matching method can accurately locate the near-duplicate fragments between the queried video and the target video. Experiments on two public datasets present that the mean average precision (MAP) of our hashing method is 0.63, which effectively improves the accuracy of video retrieval.