CAJ | 학술논문

现有的基于多示例学习的恐怖视频识别算法都是假设示例间是相互独立的，而忽略了恐怖视频中存在的上下文信息和示例包的统计特性。因此，本文提出了一种多视角融合稀疏表示模型。该模型分别从集合视角、上下文视角以及统计特性视角三个不同的视角来看待一个视频片段，并利用联合稀疏表示框架将三个不同视角融合到一个分类框架中，用来进行恐怖视频的识别。在恐怖视频库上的实验结果验证了算法在恐怖视频识别中比现有的其它算法有更好的性能和稳定性。
현유적기우다시례학습적공포시빈식별산법도시가설시례간시상호독립적，이홀략료공포시빈중존재적상하문신식화시례포적통계특성。인차，본문제출료일충다시각융합희소표시모형。해모형분별종집합시각、상하문시각이급통계특성시각삼개불동적시각래간대일개시빈편단，병이용연합희소표시광가장삼개불동시각융합도일개분류광가중，용래진행공포시빈적식별。재공포시빈고상적실험결과험증료산법재공포시빈식별중비현유적기타산법유경호적성능화은정성。
Along with the ever-growing Web ,horror videos sharing in the Internet has threatened children ’ s psychological health .It is necessary to effectively recognize and filter out these horror videos .So far ,several horror video recognition methods based on Multi-Instance Learning (MIL ) have been proposed .However ,all these methods suppose that the instances in a bag are in-dependent ,ignoring the contextual cue and statistical cue in horror videos .In this paper ,we propose a novel multi-view joint sparse coding model for horror video recognition .This model considers video from three different viewpoints including set view ,contextual view and statistical view .The set view treats a video as a set of independent frames .The context view models the contextual rela-tionship among key frames in a video using an e-graph .The statistical view represents a video as a histogram feature based on bag-of-words model .Then ,three kernel functions are designed for the three viewpoints ,respectively .Finally ,the three kernels are inte-grated into a unified multi-view joint sparse coding classification framework to recognize the horror video scenes based on recon-struction residual .Experiments on a horror video dataset demonstrate that our method’s performance is superior to the other existing algorithms .