Offline handwriten Document Regonition System for Kannada Characters

Abstract

Handwriting recognition is the ability of a machine to receive and interpret handwritten input newlinefrom multiple sources like paper documents, photographs, touch screen devices, etc. newlineRecognition of handwritten and machine characters is an emerging area of research and finds newlineextensive applications in banks, offices, and industries. The ability to detect and interpret newlinecharacters taken from an image source is known as handwritten character recognition (HCR). newlineAlthough unconstrained Handwritten Text Recognition (HTR) is an open challenge in research, newlineit has attracted researchers with the proliferation of tablet and Smartphone devices for newlinehandwriting potentially. Though the existing models on HTR majorly focus on Latin, Chinese, newlineKorean, and Japanese scripts, it has been observed that a significant amount of effort is put newlineinto the recognition of printed letters in Indian languages. Therefore, the proposed research newlineinvestigates the challenges in building an offline HCR system capable of efficiently recognising newlineKannada handwritten characters present in the document, with the ultimate goal of developing newlinea system that can be generalised for various Indian language scripts. newlineIn the present research, a handwritten character recognition model entitled UFSMSVM newlinewas proposed that consists of five major phases: image collection, segmentation, newlinefeature extraction, feature selection, and classification. The multilayer image recognition was newlineproposed using the CNN-RNN model. Long-term memory (LSTM) was created to solve the newlineproblem of vanishing gradients, and gated recurrent units (GRUs) were proposed as a simpler newlinealternative to LSTM. The efficiency of the proposed method was tested by simulating the newlinehandwritten character recognition model using the MATLAB environment followed by newlinecomparing the effectiveness of the model with benchmark models like FLM-FFNN, DIGI Net newlinemodel, and context-aware model on chars74K and MADbase digits datasets. newline