Discrete representation learning for handwritten text recognition

Abstract Handwritten text recognition, i.e., the conversion of scanned handwritten documents into machine-readable text, is a complex exercise due to variability and complexity handwriting. A common approach in recognition consists feature extraction step followed by recognizer. In this paper, we propose novel DNN architecture for that extracts discrete representation from input text-line image. The proposed model constructed an encoder–decoder network with added quantization layer which applies dictionary representative vectors discretize latent variables. parameters are trained jointly through k -means algorithm back propagation, respectively. performance suggested evaluated conducting extensive experiments on five datasets, analyzing effect handwriting recognition. results demonstrate use discretization improves deep models when compared conventional continuous representation. Specifically, character error rate decreased $$22\%$$ 22 % $$21.1\%$$ 21.1 IAM ICFHR18