GANN: a Graph Alignment Neural Network for Video Partial Copy Detection

EasyChair Preprint no. 5334

Abstract

With the advent of we-media era, massive videos have been uploaded by users to the Internet. Such a large volume of video data brings us various information. It, however, contains some fake information created by partial copy videos, which constitute infringement act and are harmful to original authors and common users. In this paper, we propose a graph alignment neural network (GANN) for partial copy videos detection. Through building a graph neural network based on video frame-level feature extracted by a pretrained convolutional neural network and their relationship, GANN automatically integrates the global representation of a video, and learns the intra-similarity between original and copied videos, and the inter-discriminative from other videos by the self-attention and cross-attention mechanism in the graph neural network. We perform experiments on the challenging dataset VCDB, which includes a variety of complex transformations in the real scene. Results demonstrate that our GANN has better detection performance than baseline methods, where the precision of GANN is close to 80%, and the recall rate reaches 65%.

Keyphrases: attention embedding, global information, Graph Neural Network, video copy detection

@Booklet{EasyChair:5334,
  author = {Xiyue Liu and Xin Feng and Pan Pan},
  title = {GANN: a Graph Alignment Neural Network for Video Partial Copy Detection},
  howpublished = {EasyChair Preprint no. 5334},

  year = {EasyChair, 2021}}