Cursive Handwritten Text Recognition using Bi-Directional LSTMs: A Case Study on Urdu Handwriting

Abstract

Recognition of cursive handwritten text is a complex problem due challenges like context sensitive character shapes, non-uniform inter and intra word spacings, complex positioning of dots and diacritics and very low inter class variation among certain classes. This paper presents an effective technique for recognition of cursive handwritten text using Urdu as a case study (though findings can be generalized to other cursive scripts as well). We present an analytical approach based on implicit character segmentation where convolutional neural networks (CNNs) are employed as feature extractors while classification is carried out using a bi-directional Long-Short-Term Memory (LSTM) network. The proposed technique is validated on a dataset of 6000 unique handwritten text lines reporting promising character recognition rates.