基于多模态的检测方法是过滤成人视频的有效手段,然而现有方法中缺乏准确的音频语义表示方法。因此本文提出融合音频单词与视觉特征的成人视频检测方法。先提出基于周期性的能量包络单元(简称EE)分割算法,将音频流准确地分割为EE的序列;再提出基于EE和BoW(Bag-of-Words)的音频语义表示方法,将EE的特征描述为音频单词的出现概率;采用复合加权方法融合音频单词与视觉特征的检测结果;还提出基于周期性的成人视频判别算法,与基于周期性的EE分割算法前后配合,以充分利用周期性进行检测。实验结果表明,与基于视觉特征的方法相比,本文方法显著提高了检测性能。当误检率为9.76%时,检出率可达94.44%。

Multi-modality based adult video detection is an effective approach for filtering pornographic information.However,existing methods lack accurate representation methods of audio semantics.Therefore,a novel method is presented in this paper to fuse audio-words with visual features for adult video detection.First,we propose a periodicity-based segmentation algorithm of units of energy envelope (EE).Audio streams are divided into sequences of EE.Second,audio semantics representation method based on EE and BoW (Bag-of-Words) is presented to describe the features of the EE as the occurrence probabilities of audio-words.Integrated weighting methods are used to fuse the detection results of audio-words and visual features.Furthermore,we propose a periodicity-based decision algorithm to judge adult videos to cooperate with the preceding periodicity-based segmentation algorithm.Therefore,we make full use of the periodicity.Our experiments show that our approach remarkably improves the detection performance compared with the method based on visual features.The true positive rate achieves 94.44% while the false positive rate is 9.76%.