Comparing Regularization Techniques for Overfitting in CNN-Based Music Genre Recognition

A key component of music information retrieval (MIR) is music genre classification, which has several uses, including automatic labelling and recommendation systems found on numerous music applications. Considering previous studies that exhibited overfitting, this study compares three widely used regularization techniques—Dropout, L2-Regularization, and Batch Normalization—to evaluate their effectiveness in minimizing overfitting. The GTZAN dataset was used, where the raw audio waveforms were converted into log-scaled Mel-spectrograms, which were used as the input to the convolutional neural network (CNN) model. A baseline CNN model and three regularized variants were trained and evaluated based on classification test accuracy, the generalization gap, and the generalization ratio. The results indicate that Batch Normalization was the most effective strategy, achieving the highest test accuracy of 75.00%. In contrast, L2-Regularization provided moderate improvement compared to the baseline model. At the same time, using Dropout by itself made the model worse at handling new songs, because randomly removing neurons interfered with learning stable patterns in how the frequency changes over time in spectrograms made from the original audio data. These results indicate that the effectiveness of regularization methods in audio-based CNNs depends strongly on the specific technique employed, with Batch Normalization offering the most promising approach for enhancing generalization in CNN-based music classification systems.