HMM-based Writer Identification in Music Score Documents without Staff-Line Removal

Writer identification from musical score documents is a challenging task due to its inherent problem of overlapping of musical symbols with staff lines. Most of the existing works in the literature of writer identification in musical score documents were performed after a preprocessing stage of staff lines removal. In this paper we propose a novel writer identification framework in musical documents without removing staff lines from documents. In our approach, Hidden Markov Model has been used to model the writing style of the writers without removing staff lines. The sliding window features are extracted from musical score lines and they are used to build writer specific HMM models. Given a query musical sheet, writer specific confidence for each musical line is returned by each writer specific model using a loglikelihood score. Next, a loglikelihood score in page level is computed by weighted combination of these scores from the corresponding line images of the page. A novel Factor Analysis based feature selection technique is applied in sliding window features to reduce the noise appearing from staff lines which proves efficiency in writer identification performance.In our framework we have also proposed a novel score line detection approach in musical sheet using HMM. The experiment has been performed in CVC-MUSCIMA dataset and the results obtained that the proposed approach is efficient for score line detection and writer identification without removing staff lines. To get the idea of computation time of our method, detail analysis of execution time is also provided.