A study on the real-time music genre classification model using convolutional neural network-based genre subclass models

초록

In this paper, we propose an N-genre music classification model and ensemble models that can more accurately classify 10 music genres based on the proposed N-genre model. The GTZAN data set is used as the music data set, and different MFCC-based color maps are compared to identify the one that produces the best classification accuracy. Two methods of model learning are used: (i) using a 30 s long music image and (ii) using a 5 s long music image, wherein the latter utilizes the voting method. The classification accuracy of these two methods is compared. Among the various types of MFCC-based color maps, the gist-rainbow method with better classification accuracy is chosen, and the 3-genre classification model is actually selected among several N-genre classification models. Furthermore, ensemble models overlapped for four genres is proposed and their classification accuracies are compared. Ensemble Model II, using the voting method, obtains an accuracy of 92.0% for the GTZAN data set. The classification accuracy of the proposed ensemble model II is compared with the those of the other reported models. © ICROS 2021.

키워드