pierrecolombo / musical-score-recognition

Personnal-Projects

##Traitement d'une partition

Introduction

Development of a Musical Recognition software in Matlab and C++ able to extract the musical data out of any kind of picture of a musical score. Signal and Image Processing techniques were widely used to isolate symbols, and Machine Learningand Statistics were used to differentiate between different symbol types.

Our Strategy

Example on this score : To achieve our goal we do the following steps :

Pre-processing the score

• if there exists a bias we apply a rotation to the picture to make it perfectly horizontal (Done using a convolution product between two halfs of the picture )
• we get the lines of the musical score by summing all the lines of matrix the score

• we get the line thickness and the line spacing by using histograms

Spliting the score into sub-score

• by using geometrical considerations we split the score into sub-score

• we substract carrefully the lines to the score to avoid noise in our recognition

Isolating into box every single symbol

• we use growth algorithms with origin the queue of the note to put every note into box

Getting the note value & the rythm

• with template matching applied in every box we decide if said box is a quaver, a semiquaver etc.... and get the pitch : (C,D,E etc...)

Conclusion : Results

Our algorithm is quite efficient on standard scores (like you have in the folder), to improve it we can give several ideas :

• use a neural network to get better accuracy, moreover template matching is very sensitive to the high variability of the symbols...
• continue to define new responsives templates to discover all "classical symbols".

musical-score-recognition's Issues

Picture one