Technology of handwriting recognition

Introducing Sanjütz’s handwriting recognition technology.

Locus of writing

Sanjütz analyzes and records the direction numbers of the handwriting locus by the user (“7,0,1,2,4,5,1,2” in the above figure). The continuity of the numbers that represent this direction tells where the locus is going turn, or is going straight, or changes the its direction.

Along these numbers, Sanjütz record the change point where the continuity is not continuity. This analysis process is done in the background of the user writing.

Information packing

When the user finishes writing, Sanjütz packs the locus numbers and some informations, such as the body size, start position, and end position

In addition, it removes redundant parts of the information and finally take hashes for dictionary references.

Recognition by dictionary

Sanjütz refer to the dictionary based on the hashed information. The dictionary has about 60,000 locus, which makes it possible to handle various stroke, glyphs.

Machine learning

Sanjütz does not use machine learning such as deep learning. The reason is:

At the beginning of development (2016), there was no library available for Apple Watch
Apple’s CoreML has been supported since WatchOS 4.0 (2017).

• It takes time to recognize
  Sanjütz wanted to be comfortable with Apple Watch 1st. generation (which I owned at the time), but it dosen’t have enough CPU power, so CoreML couldn’t reach the target recognition time.

As both Apple Watch and WatchOS evolve, we’d like to incorporate new technologies, including CoreML, in future Sanjütz updates.