State-of-the-art Optical Music Recognition system often fails to process dense and damaged music scores, where many symbols can present complex segmentation problems. We propose to resolve these segmentation problems by using a CNN-based detector trained with few manually annotated data. A data augmentation bootstrapping method is used to accurately train a deep learning model to do the localization and classification of an accidental symbol associated with a note head, or the note head if there is no accidental. Using 5-fold cross-validation, we obtain an average of 98.5% localization with an IoU score over 0.5 and a classification accuracy of 99.2%.