利用AAC壓縮域特徵之古典樂翻奏曲檢索系統

摘 要

由於網際網路及多媒體壓縮技術已相當成熟，人們對網路的需求也日益劇增，透過網路下載或分享影音資料已成為人們生活中的一部分，而龐大的音樂資料庫是很常見的，因此如何在資料庫中快速檢索出使用者所需的資料是個重要的課題。常見的搜尋引擎大多藉由文字作為輸入，但卻有標記錯誤或模糊造成檢索結果錯誤的缺點，此情況於檢索古典樂時比流行樂更常發生。

本論文針對古典音樂資料庫，利用AAC壓縮域的特徵，部分解碼出改良式離散餘弦係數，可節省約70%的解碼運算複雜度，且對係數能量作前置處理以提升準確率，將係數重新定義於十二平均律音名，並利用內積計算求得樂曲相似度矩陣，藉由尋找最佳相似度累計路徑求得其相似度分數權重平均值，以得到最後檢索結果。實驗結果顯示，所提出之方法其檢索效能MRR值為0.96，可達97%的準確率，且與傳統基於原始域檢索的方法比較，可省下90%以上的比對時間。

Classical Music Cover Song Retrieval System utilizing AAC Domain Features

ABSTRACT

With the rapid development of Internet and multimedia compression techniques, people can easily download or share multimedia data through networks. Therefore, efficient multimedia retrieval from huge multimedia database becomes an important issue. The most common method of search engines is through textual label. However, the label created by people may be ambiguous or even with errors. The situation like this in retrieving classical music occurs more often than pop music.

In our proposed system, we focus on classical music cover song retrieval in AAC compression domain. The modified discrete cosine transform coefficients are directly used to represent 12-dimensional chroma feature without a fully decoding process, and it can save about 70% decoding complexity. We truncate MDCT coefficients with low magnitude, adjust frequency boundary dynamically, and utilize dot-product calculation to get chroma similarity matrix. We calculate the similarity weighted arithmetic mean value between the songs by finding optimal similarity accumulated path and finally get the ranking results.

The experimental results show that the proposed method can reach Precision of 97% and save over 90% matching time compared with traditional approach in the waveform domain.