Markov source model for printed music decoding

Gary E. Kopec, Philip A. Chou, David A. Maltz

TL;DR

This paper introduces a Markov source model for decoding printed music notation, extending the Document Image Decoding (DID) framework to a more complex domain than previously explored. By leveraging the duality of image synthesis and decoding, the authors demonstrate a robust approach to recognizing music, paving the way for more sophisticated music information retrieval systems.

What Problem Does This Paper Address?

The paper tackles the challenge of recognizing printed music notation, a task significantly more complex than decoding standard text. Unlike text, where character meaning is largely independent of position, musical notes derive meaning from both their shape (duration) and their vertical position on the staff (pitch). This inherent two-dimensionality, along with the rich variety of musical symbols, presents a substantial hurdle for traditional document image analysis techniques. The authors aim to develop a stochastic model that can capture these complexities within the Document Image Decoding (DID) paradigm, which views recognition as a communication process involving a source, imager, channel, and decoder.

Methods

The core of the paper is the development of a Markov source model for printed music. This model is built upon the Adobe Sonata music symbol set and a finite-state language designed to represent textual music messages. The approach leverages the duality between image synthesis and image decoding, a fundamental aspect of the DID paradigm.

Key methodological aspects include:

  • Finite-State Grammar: A grammar is defined to represent the music message language, allowing for the structured description of musical elements like clefs, key signatures, time signatures, chords, and bar marks.
  • Markov Image Source Model: This model is represented as a directed graph where states represent positions and transitions represent the generation of image segments and associated message strings. Each transition has attributes like template (image segment), message string, transition probability, and displacement vector.
  • Image Synthesis: The model is used to generate music images from message strings. This process is computationally less expensive than decoding and was instrumental in developing and testing the model.
  • Decoding (Implicit): While the paper focuses on the source model, it sets the stage for decoding by defining the relationship between messages and images. The decoding process would involve finding the most likely path through the Markov model that corresponds to an observed image.
  • Sidebearing Model: The model adapts the sidebearing model, commonly used in typography for character placement, to music symbols, accounting for their spatial relationships.

Datasets

The paper uses a subset of the Adobe Sonata music font for its symbols. The music image used for illustration and decoding examples is synthesized by the model itself, based on a defined message string. No external datasets for training or evaluation are explicitly mentioned; the focus is on the model's design and its ability to handle a specific subset of music notation.

Key Insights and Ideas

  • Music as a Complex Testbed for DID: The two-dimensional nature of music notation makes it an excellent, albeit challenging, domain for testing the generality of finite-state image source models.
  • Duality of Synthesis and Decoding: The paper highlights how image synthesis can be a powerful tool for developing and validating image decoding models, especially when decoding is computationally more intensive.
  • Finite-State Limitations: The authors acknowledge that certain musical constructs, like beams and slurs, are difficult to represent within a strict finite-state framework due to their complex dependencies on multiple elements.
  • Pedagogical Value: The development of this music model served as an illustrative and pedagogical application of DID, demonstrating its applicability to domains beyond simple formatted text.

Key Results

The paper successfully develops a Markov source model capable of representing a restricted subset of printed music notation. This model handles elements such as clefs, key signatures, time signatures, chords, and bar marks. The authors demonstrate the model's capability by synthesizing a music image and then successfully decoding a degraded version of that image, producing a decoded message and a reconstructed image that closely resembles the original. While quantitative results like decoding accuracy are not the primary focus, the successful decoding example validates the model's effectiveness for the chosen subset of notation.

Limitations and Future Work

The authors explicitly acknowledge several limitations:

  • Restricted Subset of Notation: The model handles only a subset of printed music notation, excluding complex elements like beams and slurs due to their difficulty in representation within a finite-state framework.
  • Single-Voice Scores: The current language is limited to single-voice scores, although it supports intervals and chords within that voice.
  • Template Disjointness: The model sometimes violates the template disjointness constraint (where image segments from different branches should not overlap) to simplify the network, though this is noted to not cause significant decoding problems due to small overlaps.

Future work could involve:

  • Extending to More Complex Notation: Incorporating mechanisms to handle beams, slurs, and multi-voice scores.
  • Improving Grammatical Constraints: Using additional grammatical constraints to reduce decoding errors, such as the insertion of accidentals.
  • Developing a Full Decoder: Building upon the source model to create a complete decoding system for scanned music images.

Summary updated

BibTeX
@inproceedings{kopec1995markov,
  author = {Gary E. Kopec and Philip A. Chou and David A. Maltz},
  title = {{Markov source model for printed music decoding}},
  booktitle = {Document Recognition},
  year = {1995},
  doi = {10.1117/12.205814},
}

Loading related content...

Figure