Statistical features extraction for character recognition using recurrent neural network

© 2018 Pakistan Journal of Statistics. Recent studies show that recurrent neural network provided promising results for character recognition. We have extracted number of features using sliding window approach from normalized Urdu Nasta'liq text line image. The text line is scanned from right to left and top to bottom by considering Urdu script properties and extracted geometrical or statistical features, zoning and raw pixels features. We conduct four studies like sliding window with non-overlapped frame, sliding window with overlapped area with previous frame, multiple zones in a frame and raw pixels. In this paper, we evaluated MDSLTM with CTC output layer on UPTI dataset for Urdu character recognition.