Optical score recognition is a problem that is part of the Optical Music Recognition (OMR) domain. The variety of written elements present in a score, including clef, time signature, tempo, dynamics, articulation, repetitions, in addition to notes, duration and accidentals, turn this OMR task an interesting optical recognition an interesting problem in the domain of artificial intelligence. This work addresses this difficult problem using deep learning techniques and contextualizing it in the learning process of harmony students, who perform their exercises in handwritten form on paper. The preliminary results obtained, showing success rates of about 95% in the correct classification of the elements of the score in the training phase, make us envision the possibility of creating a useful tool for both teachers and students of harmony.