This project was developed for the Machine Learning class at University of Central Arkansas.

In this project, the first step in the handwritten music symbol recognition was addressed: recognize 32 isolated different musical symbols with a good grade of accuracy (at least more than 90%). In order to do that, it was necessary to get a dataset, preprocess the data, and produce a machine learning model. The dataset used was HOMUS and the method to generate the machine learning model was a Convolutional Neural Network (CNN) that includes 5 layers of convolution+ReLU (using convolutional filter 64, 128, 256, 128 and 64) and Max Pooling and the Fully Connected Layer using 1024 units/neurons in one layer and 32 units/neurons in the second layer (one for each music symbol).

The CNN model was trained with the 80% of the dataset and evaluated with the 20%. Once the model was produced, it was evaluated with self generated images (150×150 pixels), giving promising results that might be used in future research.

The entire project was developed using Python and the following libraries: numpy, cv2, random, tqdm, tflearn and matplotlib.