In this paper, we evaluated the recognition performance of BLSTM (Bidirectional LSTM) and MDLSTM (two-dimensional LSTM) neural network architecture on printed documents. We also compare the performance of 2 architectures with tesseract on same test bed. We demonstrate our experimentation on 7 Indian languages i.e. Hindi, Marathi, Tamil, Kannada, Malayalam, Bangla and Gurumukhi. The input to both the architecture will be segmented lines. The data-set used contains approximate 5000 pages for each language which then divided into train, validation and test set. The Histogram of Gradients are extracted at line level to feed into the BLSTM network. Whereas MDLSTM processes 2D image (raw pixels) of each line. The level and number of hidden layers in both the architectures are empirically selected and kept same for all the languages. The output CTC layer will contain the number of unicode present in the evaluated languages and one blank label. The input layer was fully connected to hidden layers, and these were fully connected to themselves and to the output layer. The validated result shows MDLSTM outperforms both BLSTM and tesseract for all the languages included in our experimentation.

The most important and crucial tasks in online/ offline handwritten document recognition is line/word segmentation. As compared to a printed document, line/word segmentation of handwritten document is a complicated task. The typical handwritten document have irregular skews, overlapped lines, variable gaps between lines and different size words. The recent improvements in machine learning algorithms introduced an end-to-end line level recognition of printed and handwritten text with good performance. But still line segmentation of paragraph is required before proceeding for recognition. In this paper, we proposed an improved piece-wise projection based line segmentation method for handwritten documents which is more accurate than existing methods without compromising the execution speed. Our novel methodology applies signal approximation (using fourier series in trigonometric form) and statistical approach for better line segmentation. The proposed method is capable of segmenting lines, independent of language, with performance of 99.53% on in-house CDAC dataset (having 5974 lines) and 98.11% on ICDAR Competition 2013 dataset (having 2649 lines). The dataset used for experiments consists of english, spanish, hindi and bangla handwritten paragraphs.