Enhancing Handwritten Text Identification through a Hybrid CNN-RNN Method

This research aims at the advancement of offline handwritten text recognition for moving towards a paperless environment. Our solution seamlessly unifies the power of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) architectures, using their strength for feature extraction and sequence modeling. Training and Testing: IAM dataset is used to train and test the system, which includes 87,292 training images and 4316 testing images On the system side, it will be also embodied to enrich its ability by picking up five features from a dataset (possibly designed for incorporating both spatial and sequential nature of handwritten text). We use CNN and RNN model classifiers for the classification. CNNs are widely used for extracting features from images, which is necessary to recognize handwritten characters and patterns. This usually does not work well with RNNs, but since they are designed to capture the temporal dependencies in sequential data (such as handwritten text), this is where such a model can be very successful. The results presented in the paper show how each classifier performs. Overall, the experiments in [45] demonstrate that by recognizing statement level information utilizing an RNN classifier outperforms recognition accuracy over a CNN classifier. Overall, this shows that using RNNs for handwritten text recognition can work well on the limited and messy HTR data with different writing styles. In summary, the hybrid methodology proposed in this work helps to push forward the field of handwriting without bound recognition and supports a faster succession toward digital document processing related tasks with less or even none need for manual intervention.