Naturalness learning and its application to the synthesis of handwritten characters
Ján Dolinský · Kyushu University Institutional Repository (QIR) (Kyushu University) · 2008
The modeling of human-like behaviour has recently become important in various fields of engineering.Naturalness is defined in this study as the difference between target human-like behaviour and the behaviour of an original basic system which resembles the desired target human-like behaviour but lacks human idiosyncrasy.The added value contributed by naturalness is easily understood by comparing the following: Motion trajectories of industrial robots vs. those of AIBO robots; motions in technical simulations vs. those of computer generated humans in games and movies; understandable synthesized speech vs. emotional speech; technically correct musical performances that follow the score vs. those performed by an expressive musician; letters based on font shapes vs. handwriting.All the above examples can be understood as cases where naturalness is added to a basic system.In the handwriting example, the basic system is comprised of the strokes of an original font character and the naturalness of the differences between handwritten strokes and the original font strokes.If it is possible to generate the appropriate differences (naturalness) for the strokes in the font characters, then simple addition of the naturalness to the font strokes would yield handwritten characters.An intriguing question raised by the use of such an approach is: How is the naturalness related to the font strokes which represent the basic system?Additionally, is it possible to generate the right sort of naturalness?This doctoral thesis tries to provide the answers to these questions by mathematically analyzing the relationship between the font and its naturalness using canonical correlation analysis, multiple linear regression analysis, feed-forward neural networks (FFNN) with sliding windows, and recurrent neural networks (RNN).It also attempts to show that certain systems can be viewed in terms of a naturalness learning framework.n¹hþÜYKøMW n¹hnîÙ¯Èëm¹ úYFfÒYShH Snem¹húm¹n¢Â cø¢gãYh eúLÚbkÑDWBc_L OnWgoeLhþgMúnê6UoZK30% ¦WKj Kc_ SnPK em¹Ù¯ÈëghþW_Õ©óÈ¹È íü¯Å1 húÕ©óÈhKøMWhnî n¢Âo^ ÚbgBFh¨ßU ]Sg feed-forward neural networksFFNN FFNN with sliding windown^ÚbâÇë cfeúy'nfÒLDãW_ nP íB ¶neLÅgBShoKc_L FFNN with sliding windowgAkSneú¢ÂfÒgMjDShL Kc_ SnPKôkattractorâÇêó°nÅ'k@îW RNN koecho-stateØMRNN)( eWffÒW_hS eú y'nâÇêó°LF~OúeFkjc_ Uk9o H ú dæËÃÈgnêñÕ£üÉÐïROL eYRNNâÇë HW 'ýØShLgM_ ,vnPK Õ©óÈWehW Õ©óÈWhKø MWhnîÙ¯ÈëúhYê6UnfÒko ^ÚbfÒK ÕL ï gB RNN ykROLeYShg'ý(k ØShLgMSh:W_ Ukê6UfÒnÜ(hWf Õ©óÈWkNNúnê6UÍØQ YShgKøMW W U¡W_Handwriting is a dialogue between the writer and his inner voice, where ideas organically take on shapes that reflect his original thoughts.Reading handwriting opens a dialogue with the writer, and a window not only into his original thoughts, but also his feelings and intentions.Handwriting, therefore, makes communication more valuable for both, reader and writer, and can not be replaced by a computer.If this work only serves to remind us of the inherent