An Improved Method for Eliminating False Matches in Content-Based Video Copy Detection System

摘要

It is really a challenging task to find the original video clip of a distorted duplicate among a large scale database efficiently since there are numerous variations between the original video and its copies. The difficulty mainly lies in handling the transformations while maintaining an acceptable efficiency. In this paper, an improved method for eliminating false matches both in pre-processing step and refinement step is proposed to obtain high efficiency as well as accuracy. SIFT feature points are quantized by dividing its orientation to eliminate the false matches obtained by visual index match and reduce the time cost of the later refinement. An improved method for spatial and temporal verification is proposed. In order to accelerate the frame-level verification, we concatenate weak geometric consistent algorithm and traditional 4D Hough Transform methods. And in temporary verification, a 5D Hough transform is adopted in temporal images grouping, which estimates the consistent transformation parameters from different image groups and further improves the accuracy of this detection system. We evaluate our system on the dataset of TRECVID2011 Copy Detection subject, and gains a promising performance which brings a 0.1 improvement of the NDCR measure.