CNN-RNN Hybrid model based Hindi Character Recognition

In recent years, deep learning models have tended to favor modern architectures based on the availability of huge data sets and increasing success rates. Due to the inherent complexities of the scripts and the freestyle nature of handwriting, it is extremely difficult to develop reliable lexicon-free handwritten text recognizers for Hindi. In the work that has been proposed, a CNN-RNN hybrid architecture has been used. This architecture draws its inspiration from recent developments in the use of residual blocks for training convolutional layers. A spatial transformer layer has also been included to learn a model that is invariant to handwriting distortions. The Indic Word Database is used for training and testing the model, and the results are analyzed based on word error rates and character error rates. For the purpose of script recognition, the Hindi script is taken into consideration. In the systems that have been proposed, using the synthetic data, an absolute extra reduction in WER is found.