A study on top-down word image generation for handwritten word recognition
Eiki Ishidera, Daisuke Nishiwaki · 2005
This paper describes a top-down word image generationmodel for holistic handwritten word recognition. To generatea word image, it uses likelihoods based, respectively,on a linguistic model, a segmentation model, and a charactergeneration model. In the recognition process withrespect to a given input image, it first generates, for eachword in a dictionary of possible words, a word image thatapproximates as closely as possible the input image. Themodel next calculates distance values between each generatedword image and the input image and selects for recognitionthat generated word image having the smallest distancevalue. The proposed method has been evaluated inan experiment using handwritten word images, and resultsshow it to be effective for use in handwritten word imagerecognition.