Handwritten Text Recognition Using Word Beam Search

The automated location of and transcription of handwritten text continue to be difficult problems despite several attempts. The accuracy of the following recognition process is impacted by the frequent faults in text detection and segmentation algorithms. For applications requiring sequence recognition, such as handwritten text recognition, recurrent neural networks (RNNs) are employed. When RNN is trained with CTC decoding algorithm,every output probabilities is stored in a matrix.Final text is converted using Word Beam Search algorithm. One such strategy is token passing, which can restrict the recognised text to a set of dictionary words. Because of its foundation in the Image to Sequence architecture, it can accurately sequence text that has been extracted from an image without imposing any limitations on the orientation, layout, or size of text and non-text. The proposed method outperforms token passing, best path decoding, and standard beam search decoding on the IAM datasets. There is an open-source version included. There are a variety of applications where processing vast amounts of handwritten data is important, such as document analysis, envelope address and postcode recognition, bank check amount interpretation, document analysis, and signature verification, can benefit greatly from automatic handwritten text recognition.