Translated Abstract
Compared to the traditional music computer music can be safe, long-term stored, with more unique and rich artistic expression, which can also be shared easily and quickly by people through the internet. Therefore, the combination of computer and music is the inevitable trend of technological development, and high-speed information transmission, more and more powerful hardware functions create a better condition for the computer development in the music field. In this context, the importance of music emotion classification becomes more prominent in music retrieval and music recommendation.
Music emotion analysis can be divided into three parts: music characteristic model, music emotion model and classification cognition model. Traditional music emotion classification methods can not always take account of the subjectivity, hierarchy and athleticism of music emotion all at the same time. This thesis chooses the MDI file as the research object, and proposes an improved music emotion classification method based on neural network. This method introduces the concept of ＂music energy＂, and makes use of the melody area of music to calculate the similarity between the sections of music, so as to realize the segmentation of music in time dimension. Experiments show that the segmentation performance is the best when the cosine similarity threshold sets 0.5.In this thesis, we use Hevner emotional ring model, and improve the methods to meet the hierarchy and movement characteristics of music emotions. BP neural network is selected as the classification cognition model, and a personalized training method is proposed to realize the personalized emotion classification for different users. Experiments show that the music emotion classification method based on neural network proposed in this thesis achieves an average correct rate of 83% for music within 180 seconds, which is better than traditional algorithm.
Finally, this thesis designs and implements a music retrieval system based on emotion analysis. Based on the personalized neural network training method proposed in this thesis, this system can classify music emotion for different users, so as to meet the needs of users＇ music retrieval. This thesis describes the construction process of the system from three aspects: needs analysis, design and implementation, and proves the practicability of the system by testing.
Corresponding authors email
