Abstract
In this paper, we propose an approach along with an extended ablation study to address the writer identification task in handwritten scores. To benefit from machine learning methods and train them when working with these types of images, traditional approaches tend to apply some standard transformations, such as reshaping the image or randomly choosing a small image patch. This process can seriously degrade the information received by the model and therefore, it risks learning non-discriminatory information. To address these problems, we propose a sliding tile-based approach for the task of writer identification in music scores in two stages. The first stage benefits from symbol detection for the sole purpose of identifying optimal regions containing musical information. And a second stage uses this music content information to compute the final classification at the full-page. We present an ablation study together with a new database containing musical scores extracted from the music department of the National Library of France - BnF named REMDM Autographs. We present the results of our explorations for two databases, the new corpus, and the well-known public CVC-MUSCIMA database. When comparing the performance of the tile approach versus the full-page approach, we see an undeniable performance improvement of more than 45% in both databases.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
