A Feature Extraction from Compressed Multimedia Data For Content-based Multimedia Data Retrieval

多媒体数据库（Multimedia Database，MDB）是管理大量多媒体数据的基础和重要系统。要建立一个有效的多媒体数据库需要解决两个问题：一是如何有效地存储多媒体数据，二是如何有效地检索多媒体数据。为了解决这些问题，我们提出了一个简单而清晰的框架，基于内容的多媒体数据检索系统使用MPEG 4域，并提供了一个新的音乐特征，基于内容的音乐检索的新框架。基于内容的多媒体数据检索系统的新框架与传统的基于内容的多媒体数据检索系统框架的一个重要区别在于特征提取方法的不同。在新的框架上，从压缩数据或压缩方法中提取用于描述多媒体数据的内容的特征，而在典型的框架上，从行数据（例如，原始图像数据、音乐的信号数据等）中提取特征。在此框架下，我们针对基于内容的音乐检索系统进行了研究，并实现了一个原型系统， 关于我们 允许我们以比特率不同的压缩音乐数据为例进行查询。为了实现一个基于内容的音乐检索系统，本文提出了一种新的音乐特征，它满足以下要求：1。可从MPEG-4/音频域计算，2.对于比特率是鲁棒的，3.从理论上证明了比特率为B_1的第i个自相关系数可以近似比特率为B_2的第j个自相关系数，其中j=[（B_1）/（B_2）i]。因此，我们将自相关系数与比特率的关系称为与比特率的近似关系。通过实验验证了从MPEG 4/音频域提取的自相关系数确实满足与比特率的近似关系，并利用这种近似关系提出了一种新的音乐特征，并在音乐检索中进行了实验，评估了不同比特率下的音乐特征的检索性能。实验结果表明，该音乐特征具有良好的检索性能，对不同比特率的查询具有良好的检索性能。少

Multimedia Database(MDB)is a fundamental and important system to manage large set of multimedia data. There are two problems to make an effective MDB.One is how to efficiently store multimedia data, another is how to effectively retrieve multimedia. For solving these problems, we propose a simple and clear framework for a content-based multimedia data retrieval system using MPEG4 domain and provide a new music feature for a content-based music retrieval on the new framework. A big difference between the new framework of a content-based multimedia data retrieval system and a typical framework of that is a feature extraction method. On the new framework, a feature for describing a content of multimedia data is extracted from the compressed data or a compression method, while on a typical framework a feature is extracted from row data(e.g.raw image data, signal data of music, etc.).We target on a content-based music retrieval system on the new framework and make a prototype system which a … More llows us queries by example of the compressed music data whose bit rate are various. For making a content-based music retrieval system, we provide a new music feature which satisfies the following requirements ;1. computable from MPEG-4/audio domain,2. robust for bit rate,3. computable from the compressed data without decoding.We show theoretically that the i-th autocorrelation coefficient with bit rate B_1 can approximate the j-th autocorrelation coefficient with bit rate B_2, where j=[(B_1)/(B_2)i]. So, we call a relation of autocorrelation coefficients with respect to bit rates an approximation relation with bit rate. And we experimentally confirm that the extracted autocorrelation coefficients from MPEG4/audio domain actually satisfy the approximation relation with bit rate.Then, we provide a new music feature by using the approximation relation with bit rate, and experiment on the music retrievals to evaluate the performance of retrievals by using the music feature with various bit rates. The results of experiments show a good performance of retrievals, and that the proposed music feature have a good performance of retrieval for queries with various bit rate. Less