Handwritten text recognition using deep learning method

Offline handwritten text recognition has been widely used in the last decades, especially in several fields like historical document cognition. Deep learning has the power to address the problem of digitizing traditional writing. Each algorithm in deep learning can handle a specific task or solve a particular problem. In this thesis, we use a convolutional neural network for extracting distinct features of the characters in input words. Combining characters needs a neural network that can find combinations in a given sequence since multiple characters can form a specific word. A recurrent neural network can do this. The long sequences need memory in a neural network to remember the combination. This is the task of long-short-term memory (LSTM) built a hybrid deep neural network which contains three layers of convolutional neural network (CNN) followed by bidirectional long-short term memory (BLSTM) to encode the input image by giving a matrix of probabilities over characters at each image position. Connectionist Temporal Classification (CTC) operation uses this matrix to find the most suitable match between predicted and target output for computing the loss function. We tried different parameters value in the model till we obtained the best ones. We evaluated our model on the IAM dataset and got a reasonable error rate.