PAKPAHAN, KORNELIUS BOAZ APRILIO (2025) A COMPARATIVE ANALYSIS OF PRE- PROCESSING TECHNIQUES FOR OPTICAL MUSIC RECOGNITION: ASSESSING THE EFFICACY OF STAFFLINE REMOVAL. S1 thesis, UNIVERSITAS KATOLIK SOEGIJAPRANATA SEMARANG.

Preview	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-COVER_a.pdf Download (860kB) | Preview
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-BAB_I_a.pdf Restricted to Registered users only Download (803kB)
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-BAB_II_a.pdf Restricted to Registered users only Download (696kB)
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-BAB_III_a.pdf Restricted to Registered users only Download (1MB)
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-BAB_IV_a.pdf Restricted to Registered users only Download (1MB)
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-BAB_V_a.pdf Restricted to Registered users only Download (726kB)
Preview	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-DAPUS_a.pdf Download (721kB) | Preview
	Text 21.K1.0038-KORNELIUS BOAZ APRILIO PAKPAHAN-LAMP_a.pdf Restricted to Registered users only Download (1MB)

Abstract

Standard Optical Music Recognition (OMR) systems use staff-line removal as part of the pre-processing step to improve detection accuracy for the symbols. However, with the emergence of deep learning, this step started to seem somewhat unnecessary. In this study, we present a comparative study of three pre-processing scenarios, namely, staff-line removal via Projection Profile Analysis (PPA) subject to morphological operations (morphological), RunLength Encoding (RLE), and no staff-lines removal. In order to assess its effects on recognition performance using Symbol Error Rate (SER), a CNN+LSTM model was trained on the PrIMuS dataset—a collection of 87,678 high-quality monophonic score images—under each of these circumstances. In this study we found that the scenario without removing staff lines leads to the lowest validation SER (0.62) compared with PPA+Morphological (SER: 0.75) and RLE scenarios (SER: 0.82). Although staff lines removal approaches such as PPA (0.01s/sample) and RLE (0.1s/sample) decreased computational costs, they were associated with artifacts resulting in more degraded models. The results make a case for prioritizing symbol integrity over pre-processing efficiency in deep learning-based OMR systems, and may help mitigate pre-processing issues in future work towards streamlined workflows and model optimization.

Actions (login required)

View Item