Developing technology to automatically generate human cursive

The Pattern Recognition Lab has published a paper on technology that automatically creates handwritten sentences by imitating human handwriting.

The system being developed by the research team simplifies the handwritten words written in the left frame first through Skeletonization and online sequences. Then, the style to be written is applied to the previously entered string. After rendering this, the handwriting is output as an image. In this method, if you input one word written by a person’s hand, it can output a person-specific character written in another letter.

Until the transformation was performed, the machine learning algorithm cycle GAN (CycleGAN) was used. In addition, because it is impossible to resample at a constant speed in learning, it is said that the maximum acceleration sampling is performed by sampling the points on the line at intervals.

According to the research team, as a result of conducting a Turing test to determine whether the character generated by the system is human, 50.3% of the subjects can identify the handwriting printed by this method as fake. At this time, 67.7% of those who judged it as actual writing. However, the percentage that the identification system identifies false handwriting is 25%, and the general level is unclear to the general public, but the professional system cannot be deceived.

The technology is still in the development stage, and there is a problem that it inherits the bad habits of people who skip or write punctuation marks. It also revealed that there is a possibility that if you are a forensic expert, you may know that it is a different handwriting.

Of course, once this skill is completed, even people with physical problems can write sentences with their own handwriting. In addition, even if you have previously needed to scan and capture handwritten messages, you can save time. In addition, it can be helpful for handling one’s handwriting in a virtual reality world including games or creating a sentence recognition engine sample that requires a lot of training samples. In addition, stronger identification methods can be developed in forensic research. Related information can be found here .