A Systematic Framework for Sanskrit Character Recognition Using Deep Learning

Abstract

Sanskrit is widely acknowledged to be among the world’s oldest surviving classical languages, and yet its usage has continued to decline unabated in the present milieu. Such insidious erosion of popularity is directly attributable to the absence of native speakers of the language and the perceived inaccessibility of Sanskrit to contemporary audiences. Notwithstanding, the language remains historically and culturally inseparable from the subcontinent, with numerous religious manuscripts, epigraphical inscriptions, edicts and scientific literature written in the Sanskrit script. Attempts made to resuscitate the language have been largely unsuccessful as these attempts have
relied extensively on laborious human transcription and translation. Such manual endeavors can be superseded by the use of efficient computational techniques to facilitate the efficient transcription of voluminous manuscripts written in the Sanskrit script.

The emergence of deep learning frameworks has enabled researchers to overcome the draw backs of conventional machine learning algorithms in developing efficient and extensible character recognition systems. Notwithstanding, the advancement of character recognition frameworks varies across different Indic scripts.

In this context, this paper introduces an extensible framework for the transcription of hand written Sanskrit manuscripts. In the absence of a benchmark dataset of handwritten Sanskrit characters, the authors introduce a comprehensive dataset to facilitate further downstream segmentation. The dataset, on augmentation, comprises over a hundred thousand samples and has been collected from over a hundred individuals. The paper explores an integrated approach to segmentation and accordingly delineates a systematic methodology for  effectively segmenting Sanskrit words, incorporating techniques such as thresholding, zone-based classification, median bisection
and projection profiles. The proposed technique accommodates a diverse array of characters and modifiers present in the Sanskrit script. Subsequently, a concurrent deep learning architecture parallelizes transcription using Neural Networks (CNN and Residual Networks). The deep learning models show accuracies exceeding 90%. This paper attempts to benchmark the significance of
systematic approaches to machine transcription of low-resource languages.