基于情感的家庭音乐相册自动生成研究与实现

发布时间：2018-12-13 07:32

【摘要】：随着互联网以及社交网络的发展,电子相册逐渐成为应用广泛的基础服务之一,而如何提高相册的用户体验变得尤为重要。本文提出基于情感的家庭音乐相册自动生成研究,旨在解决为用户喜爱的音乐自动推荐与其情感表达相近的相册图片问题。本文从音乐和图像蕴含的情感出发,选取能描绘其情感的特征,利用典型相关算法分析音乐与图像特征在同一情感类别下的线性相关性,提出了基本映射模型并实现了音乐相册的自动生成。本文根据MIREX的情感分类标准[93],将情感主要分为五种:兴奋的,愉悦的,悲伤的,诙谐的,愤怒的。根据Thayer的二维情感模型[61],利用V(Valence)、A(Arousal)值确定音乐和图像的情感坐标,从而界定图像和音乐的情感类别范围。挖掘图像和音乐中能较好表征其情感的特征,音乐上选取MFCC和RASTA-PLP特征,图像上选取比较直观的颜色特征。在算法上使用了善于分析两组变量之间关系的典型相关算法(CCA),该算法解决了图像特征与音乐特征之间的异构问题,实现了跨模态的特征融合。在实验过程中,本文将采用CCA方法和随机推荐,人工推荐的方法进行分析和比较,结果表明,本文采用CCA方法获得的实验结果在人工评价时得到67.45%的满意度,与人工推荐的评价结果(78.09%)比较接近。在测试音乐情感已知与未知的对比下,实验结果表明采用CCA方法具有一定的有效性,本文提取的音乐特征与图像特征之间存在一定的线性相关性。
[Abstract]:With the development of Internet and social network, electronic photo album is becoming one of the basic services, and how to improve the user experience of photo album becomes more and more important. This paper presents a research on automatic generation of family music album based on emotion, which aims to solve the problem of automatic recommendation of music for users and its emotional expression. Based on the emotion contained in music and image, this paper selects the features that can describe their emotions, and analyzes the linear correlation between music and image features under the same emotion category by using canonical correlation algorithm. The basic mapping model is proposed and the automatic generation of music album is realized. According to MIREX's criteria of emotion classification [93], emotion is divided into five categories: excited, joyful, sad, witty and angry. According to Thayer's two-dimensional emotional model [61], the emotional coordinate of music and image is determined by using V (Valence), A (Arousal) value, so as to define the emotional category of image and music. Mining images and music can better represent their emotional characteristics, music selected MFCC and RASTA-PLP features, image selection of more intuitive color features. In the algorithm, a typical correlation algorithm, (CCA), which is good at analyzing the relationship between two groups of variables, is used. The algorithm solves the heterogeneous problem between image features and music features, and realizes the cross-modal feature fusion. In the course of the experiment, the CCA method, random recommendation method and artificial recommendation method are used to analyze and compare the results. The results show that the experimental results obtained by CCA method are satisfactory to 67.45% in manual evaluation. The result was close to that of manual recommendation (78.09%). By comparing the known and unknown musical emotions, the experimental results show that the CCA method is effective, and there is a certain linear correlation between the music features extracted in this paper and the image features.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP393.09

【相似文献】