基于分形特征的音频检索

Fractal Feature-based Audio Retrieval

提出利用分形几何抽取音频特征的全局化音频检索,将其学习阶段计算音频数据库中每个音频的分维作为特征向量,保存在音频特征数据库中,并建立索引。其检索阶段则首先计算查询音频的分维,然后从音频数据库中快速找出分维最相似的若干音频对象。分维刻画了音频的内在属性如自相似性,使其具有片段检索对匹配的起点不敏感、抗噪音、检索速度快等优点。用FRACTAL,MFCC和SOLAR3种方法对数据集分别检索,实验结果表明基于分维的音频检索在性能和时间复杂度上有显著优势。

The fractal geometry-based feature extraction is proposed for audio retrieval system. During the learning process, the system computes the fractal dimension as the feature vector for each audio in audio database and then saves it in the feature vector database. In the retrieval process, the fractal dimension for the query audio is firstly extracted, by which the most similar audios from the audio database are retrieved. The fractal dimension is intrinsic for each audio such as self-similarity so as to make it not sensitive to noise and position of the audio fragment to be retrieved from the long audio. It also retrieves the audios quickly. Compared with FRACTAL, MFCC and SOLAR, the experimental results validate that the proposed approach advances in performance and time complexity.