Handwriting Normalization by Zone Estimation Using HMM/ANNs
Joan Pastor-Pellicer, Salvador España-Boquera, Francisco Zamora-Martínez, María José Castro-Bleda · 2014
Offline handwritten text recognition requires several preprocessing stages. Many different preprocessing techniques have been proposed in the literature based either on geometrical heuristics or on statistical models. Unfortunately, these approaches usually fail when dealing with short sentences or isolated words. One statistical technique for text line preprocessing is based on the detection and classification of local extrema, by means of neural networks, to determine the reference lines delimiting the different areas. Unfortunately, it is not robust against a single bad classified point. This paper proposes a novel method to normalize handwritten text lines based on a supervised statistical model which takes into account all pixels instead of just the local extrema. A Hidden Markov Model hybridized with an artificial neural network is applied column-wise in order to segment each column of the handwritten line into three areas. The reference lines obtained in this way are used to normalize the image afterwards. The technique has been empirically validated on the IAM offline database.