基于BoF的高效音乐过滤指纹

为了提高在海量音乐资源库中对未知信息的音乐片段的检索效率,提出一种快速鲁棒的音乐过滤指纹。对经典的Philips音频指纹采用bag-of-features算法,生成一种数据量大幅缩小的过滤指纹,可以快速地大幅缩小检索范围,从而有效地提高检索速度。实验显示,对相同的音乐片段进行过滤检索,平均过滤时间仅为传统顺序检索时间的1/7 328,平均过滤幅度可达99.67%以上。并对MP3转换、重采样、随机剪切处理的检索召回率仍在99.47%以上,理论准确率接近100%。

An efficient music filtering fingerprint based on BoF

Aiming at improving the retrieval efficiency of unknown music clip in the massive music library, a fast and robust music filtering fingerprint is proposed. By bringing the bag-of-features algorithm to the classical Philips audio fingerprint, a filtering fingerprint is generated with the fingerprint data reduced multiply to narrow the scope of retrieval quickly and sharply, thus the retrieval efficiency is greatly increased. Experiments show that retrieving a 10 seconds audio clip with filtering fingerprint, the average filtering time is only 1/7328 of the traditional sequence retrieval time, but the filtering range is above 99.67%. After applying MP3 conversion, resampling and random shearing, the recall rate is still over 99.47%, and the theoretical accuracy is nearly 100%.