Music Classification Scheme Based on EfficientNet-B3

Several studies have been conducted music genre classification methods for music streaming services to effectively search and recommend music. The existing methods accurately classify known music genres, whereas they cannot distinguish unknown from known music genres or correctly classify unknown music genres as specific known music genres. Thus, this study proposes an unknown music genre classification (U-MGC) scheme that classifies both known and unknown music genres. The U-MGC generates mel-spectrogram images from audio data to indicate frequency changes over time. Then, U-MGC classifies the audio data into specific music genres by inputting the generated images into the EfficientNet-B3 model, which is constructed based on the placeholder for open-set recognition (PROSER) algorithm. Since the U-MGC is generalized for the entire music genre, it accurately classifies different types of unknown music genres. The evaluation results showed that the classification performance of the proposed U-MGC was 74.1% for the GTZAN dataset and 65.6% for the FMA large dataset. These U-MGC improved accuracy by 1.7% to 2.1% compared to the existing music genre classification methods.