Gabriel Vigliensoni

Creating a large-scale searchable digital collection from printed music materials

In this paper we present our work towards developing a large-scale web application for digitizing, recognizing (via optical music recognition), correcting, displaying, and searching printed music texts. We present the results of a recently completed prototype implementation of our workflow process, from document capture to presentation on the web. We discuss a number of lessons learned from this prototype. Finally, we present some open-source Web 2.0 tools developed to provide essential infrastructure components for making searchable printed music collections available online. Our hope is that these experiences and tools will help in creating next-generation globally accessible digital music libraries.

Pixel-wise binarization of musical documents with convolutional neural networks

Binarization is an important process in document analysis systems. Yet, it is quite difficult to devise a binarization method that perform successfully over a wide range of documents, especially in the case of digitized old musical manuscripts and scores with irregular lighting and source degradation. Our approach to binarization of musical documents is based on training a Convolutional Neural Network that classifies each pixel of the image as either background or foreground. Our results demonstrate that the approach is competitive with other state-of-the-art algorithms. It also illustrates the advantage of being able to adapt to any type of score by simply modifying the training set.

Staff-Line Detection on Grayscale Images with Pixel Classification

Staff-line detection and removal are important processing steps in most Optical Music Recognition systems. Traditional methods make use of heuristic strategies based on image processing techniques with binary images. However, binarization is a complex process for which it is difficult to achieve perfect results. In this paper we describe a novel staff-line detection and removal method that deals with grayscale images directly. Our approach uses supervised learning to classify each pixel of the image as symbol, staff, or background. This classification is achieved by means of Convolutional Neural Networks. The features of each pixel consist of a square window from the input image centered at the pixel to be classified. As a case of study, we performed experiments with the CVC-Muscima dataset. Our approach showed promising performance, outperforming state-of-the-art algorithms for staff-line removal.

GRAIL: Database Linking Music Metadata Across Artist, Release, and Track

Linking information from multiple music databases is important for MIR because it provides a means to determine consistency of metadata between resources/services, which can help facilitate innovative product development and research. However, as yet, no open access tools exist that persistently link and validate metadata resources at the three main entities of music data: artist, release, and track. This paper introduces an open access resource which attempts to address the issue of linking information from multiple music databases. The General Recorded Audio Identity Linker (GRAIL - api.digitalmusiclab.org) is a music metadata ID-linking API that: i) connects International Standard Recording Codes (ISRCs) to music metadata IDs from services such as MusicBrainz, Spotify, and Last.FM; ii) provides these ID linkages as a publicly available resource; iii) confirms linkage accuracy using continuous metadata crawling from music-service APIs; and iv) derives consistency values (CV) for linkages by means of a set of quantifiable criteria. To date, more than 35M tracks, 8M releases, and 900K artists from 16 services have been ingested into GRAIL. We discuss the challenges faced in past attempts to link music metadata, the methods and rationale which we adopted in order to construct GRAIL and to ensure it remains updated with validated information.

Document Analysis for Music Scores via Machine Learning

Content within musical documents not only contains musical notation but can also include text, ornaments, annotations, and editorial data. Before any attempt at automatic recognition of elements in these layers, it is necessary to perform a document analysis process to detect and classify each of its constituent parts. The obstacle for this analysis is the high heterogeneity amongst collections, which makes it difficult to propose methods that can be generalizable to a broader range of sources. In this paper we propose a data-driven document analysis framework based on machine learning, which focuses on classifying regions of interest at pixel level. The main advantage of this approach is that it can be exploited regardless of the type of document provided, as long as training data is available. Our preliminary experimentation includes a set of specific tasks that can be performed on music such as the detection of staff lines, isolation of music symbols, and the layering of the document into its elemental parts.