Description

Tyler Gutowski
1 Introduction
Music genres often reflect unique artistic, cultural, and historical elements. The visual design of album covers can be an indicator of these genres, providing a rich dataset for machine learning. This project explores the classification of music genres using visual cues from album covers with the application of Convolutional Neural Networks (CNNs). We hypothesize that CNNs can effectively classify images into music genres, offering insights into the relationship between visual art and music.
2 Data Preparation
2.1 Data Collection
2.2 Data Preprocessing
3 Methodology
3.1 Model Architecture
The architecture of the CNN was critical to the project’s success. The model consisted of multiple layers, each designed to capture different aspects of the image data. The Conv2D layers were responsible for extracting features from the images, while the MaxPooling2D layers reduced the dimensionality of these features. The Flatten layer converted the 2D features into a 1D vector, which was then processed through Dense layers for classification. Dropout layers were included to prevent overfitting.
3.2 Training Process
The model’s training involved several steps. Firstly, the preprocessed dataset was loaded into the model. The genre labels were one-hot encoded to facilitate classification. The dataset was then split into training and testing sets.
The model was compiled with a categorical cross-entropy loss function and optimized using the Adam optimizer. The training process was monitored for accuracy and loss metrics to ensure effective learning.
4 Results and Discussion
4.1 Model Evaluation
The model’s performance was evaluated using the testing set. The evaluation focused on accuracy and loss metrics, providing a quantitative measure of the model’s effectiveness. A confusion matrix was generated to visualize the model’s performance across different genres.
4.2 Visualization
Visualization tools were used to illustrate the model’s classification accuracy per genre. The confusion matrix provided insights into the genres that were most accurately classified and those that posed challenges.
5 Conclusion
This research demonstrates the potential of using visual cues from album covers for music genre classification. The CNN model showed promising results, indicating that album cover imagery contains significant information relevant to genre classification. This study opens avenues for further research in the field of music classification, particularly in exploring the relationship between audio and visual data in understanding music genres.