AN HMM-BASED OCR FRAMEWORK FOR TELUGU USING A TRANSFER LEARNING APPROACH

Abstract

Optical character recognition (OCR) for complex scripts such as Telugu has gained much attention over the past decade due to the significant advancements made in this area of research. The Telugu OCR framework in this work proposes a Hidden Markov model based approach using transfer learning to estimate the emission probability parameter of the model. This approach incorporates knowledge of the Telugu language into the framework via the hidden Markov model, while the pre-trained convolutional neural network, VGG-16, aids in estimating the emission parameter. A comparative analysis of two estimation techniques for estimating the emission parameter is also provided. One method utilizes Gaussian mixture models clustering using feature vectors obtained from VGG-16 and the second method utilizes the softmax outputs from VGG-16 to obtain emission probability estimates. The results from this framework show that using a pre-trained CNN for parameter estimation instead of as a classifier significantly reduces the resources required for developing an OCR framework for Telugu compared to implementing a CNN framework from scratch.