Re-exploring the structure of Chinese character images

Staff Picks
, , , , , , , ,
1
◔
11 Likes
2

news-congrads-kkluyshnik-02-04-19.jpg – you have earned Featured Contributor Badge 2015-11-02_10-07-11.png
Your exceptional post has been selected for our editorial column Staff Picks http://wolfr.am/StaffPicks and Your Profile is now distinguished by a Featured Contributor Badge and is displayed on the Featured Contributor Board. Thank you!

3

Just a teaser here – I plan to use the code (and explanations) in this post as a basis for submitting animated Rorschach test images for the “Computational Art Contest 2022” .

1 Like
4

Thank you for the recognition, Moderation Team!

5

Can’t wait to see it :slight_smile:

1 Like
6

Nice followup! I hope @Silvia Hao will see this. And I am also looking forward to the animated Rorschach post. :slight_smile:

3 Likes
7

This post elaborates on how to make the Rorschach animations:
“Rorschach mask animations” .

8

Other versions:

  1. WordPress blog post (HTML)

    • Made with the package M2MD.

  2. GitHub post (Markdown)

1 Like
9

Thanks a lot for pinging me here!

1 Like
10

Nice post, Anton! In terms of interpretable model, I’m always a big fan of traditional statistical analysis!

One central interest on characters and fonts (not just CJK) is the topological relationship between strokes. It’s somehow related but beyond properties in image space. There are neural net approach to this topic, like FontRNN, etc. I guess they captured those topology nicely, so I wonder how can we “extract” the information, and what can we say from applying “traditional” statistical analysis on them. If we can say something meaningful from that, then we get insight and they become interpretable deep learning models.

2 Likes
11

Thank you for your feedback, Silvia!

[…] I’m always a big fan of traditional statistical analysis!

That comment provided additional motivation to “wrap up” a first version of the post “Handwritten Arabic characters classifiers comparison” .

In that post I show how using dimension reduction and nearest neighbors we can get better classification results than using “standard” convolutional networks. Of course, we also get additional insight.

One central interest on characters and fonts (not just CJK) is the topological relationship between strokes. It’s somehow related but beyond properties in image space.

  1. In the linked post I had to crop-&-resize the images in order to get better results. That is, in some sense, related to having hard time to “discover” the topology in Arabic characters.

  2. Another approach is to actually make some cheap transformations that expose the topology. That is, of course, not that trivial.

  3. People writing Chinese characters – after years of schooling – follow certain order of the strokes. I think that is much more fruitful type of data utilization. (And as far as I know it is used in the Chinese character completions in different OS’es.)

2 Likes
12

Just curious how close your reduced set of auto radicals compared to those of the Cang Jie Input radicals?

1 Like
13

In general, the two steps:

  1. Radical finding

  2. Dimension reduction with found radicals

can be iterated multiple times and the obtained radical basis can be used to represent any Chinese character. And, yes, I think that basis is going to be similar to the Cangjie input method.

I am not sure how fruitful or useful that representation would be. It depends on the purposes for its use. For example, in general the Machine Learning (ML) algorithms that provide good explanations are not that good for, say, ML classification or approximation.

2 Likes