Image-based music genre classification using convolutional neural networks

Abstract

This dissertation presents a full investigation into the application of machine learning, specifically Convolutional Neural Networks (CNNs), in the music genre classification using vinyl album covers. Leveraging the latest advancements in deep learning and computer vision, the study introduces two models: one employing batch normalization and another utilizing Concept Whitening (CW) techniques to enhance model interpretability. The main objectives are to evaluate these models' classification accuracy and interpretability. Using robust evaluation parameters, both models exhibit good accuracy rates in classifying music genres based on vinyl album covers. Notably, the concept whitening model adds another layer of interpretability, unraveling the black-box features found in standard neural networks. Empirical findings indicate that concept whitening enhances model interpretability and competes effectively in predictive performance. This project serves in the pursuit of reliable and transparent image-based music genre classification systems. By comparing the two models on both accuracy and interpretability fronts, the study shines a light on the viability of incorporating concept whitening into standard CNN architectures for more explainable AI applications.

Date of Award	18 Oct 2023
Original language	English
Awarding Institution	Universidade Católica Portuguesa
Supervisor	Nicolò Bertani (Supervisor)