Scattering Network features based recognition on verification framework for Malayalam printed and handwritten Character Recognition

Abstract

Optical character recognition (OCR) is the process of transforming the scanned newlinedocument images into the machine editable and searchable format. In Indian languages, the research efforts conducted toward OCR systems, and the document image resources available for research are comparatively less. The objective of the research work is to implement a feature-based character recognition system for Malayalam language, one of the official languages in India. A large number of character glyph, the structural resemblance between different character shapes and the non-availability of open source language resources are the newlinemain challenges in implementing OCR system for Malayalam language script. newlineTill date, no standard character image database is available for Malayalam newlinelanguage script. The present research work builds a character-level printed newlineand handwritten character image database for Malayalam language, by which a uniform assessment of different existing methods for Malayalam character recognition can be achieved. The created character image database has 29,302 handwritten and 52,265 printed Malayalam character images. In feature-based character recognition systems, the employed feature extraction technique plays a significant role and affects the overall performance of the system. The proposed research work employs scattering convolutional network-based features for Malayalam handwritten and printed character recognition. Scattering convolutional network depends on the scattering transform which generates invariant feature descriptors with the support of wavelet decomposition and non-linear operators. The scattering feature maps are utilized in Malayalam character recognition through singular value decomposition to capture the discriminating features in higher-layers of scattering network in lower dimensions. Another technique chosen for employing the scattering network in character recognition is through integrating the scattering feature newlinemaps with the convolutional neural networks (CNN). The accuracy achievement of