|(1) Posted by Viktoras Paliulionis [Friday, Nov 11, 2022 01:45]|
A new web-based chess diagram recognition tool is available at https://helpman.komtera.lt/chessocr.
It was created by my daughter Alina as a course work at university.
|(2) Posted by James Malcom [Friday, Nov 11, 2022 03:08]|
This will be very helpful to many, I give thanks.
|(3) Posted by Hauke Reddmann [Friday, Nov 11, 2022 10:37]|
As a CS student - how does the "intelligence" of the program work?
(Chess pieces come in many forms...Obviously it shouldn't be too hard
to throw a neural net at the standard form, but what about all others?)
In particular, if the program is fed a problem of the genre
"Assign the letters to chess piece types", does it manage that? :-)
|(4) Posted by seetharaman kalyan [Friday, Nov 11, 2022 10:58]|
Just tested. It is very nice 👍
|(5) Posted by Eric Huber [Friday, Nov 11, 2022 11:28]|
I have tested it with a picture from Valery Surkov's website: https://ic.pics.livejournal.com/valsur/25495879/138264/138264_original.jpg
(the photo can be found on his post about G. Baev https://ru-chess-art.livejournal.com/577105.html)
It works just fine.
Thank you very much !
|(6) Posted by edderiofer [Friday, Nov 11, 2022 11:33]|
I've already been using the ChessVision.ai browser plugin, but being able to copy-paste from clipboard is useful.
|(7) Posted by Viktoras Paliulionis [Friday, Nov 11, 2022 15:43]|
A neural network was trained to recognize 12 standard piece classes. The training data set included diagrams of various sizes, styles, and qualities.
|(8) Posted by Jakob Leck [Monday, Nov 14, 2022 09:11]|
Nice work. As I guess from the js code the main work is done on a server, and in principle one could call the REST api from a different frontend, e.g. on a mobile device, right?
|(9) Posted by Joost de Heer [Monday, Nov 14, 2022 11:33]|
Fairy pieces aren't recognised.
|(10) Posted by Hauke Reddmann [Monday, Nov 14, 2022 18:00]|
@Joost: Neural Net. It can't recognize a simple rotation by
180°, it has no "concept" of that or mirroring. (But obviously
it would be straightforward to add those as training items.)
|(11) Posted by seetharaman kalyan [Monday, Nov 14, 2022 18:48]|
It was after all a college project
|(12) Posted by Viktoras Paliulionis [Monday, Nov 14, 2022 23:02]|
In theory, the API could be called from different frontends. However, due to limited computing resources it is not planned for wider use. On a mobile device, you can use a web browser, no other application is required.
The goal of this work was to achieve the best possible accuracy of standard piece and standard diagram (not rotated) recognition. Although the neural network could be trained for other pieces (e.g., fairy).
No more posts
MatPlus.Net Forum Internet and Computing Chess OCR