Performance of Convolutional Recurrent Networks for Handwritten Text Recognition

The globe has been swept up in the popularity of the optical character recognition technology, which can analyze written characters in high-quality photographs. OCR, however, performs poorly in real-world settings for a number of reasons, including differences in the style of writing, varied spacing, poor character detection due to inconsistent handwriting, and complexity of detection due to low image quality. Hence, it is still difficult to recognize text from handwriting. Convolutional and Recurrent neural networks are the foundation of the most effective offline handwritten text recognition techniques currently available. But, they are susceptible to data loss while processing huge text pictures because they are able to learn only higher level features of source domain. To overcome this challenge, many designs based on gated CNN have been developed which also pick up the domain agnostic features. In this study, models based on various designs are constructed and then compared. To assess whether architecture has shown impressive outcomes in recognition of handwritten text, the precision, character error rate, and word error rate are measured.