search

danpla / dpscreenocr

Program to recognize text on screen
https://danpla.github.io/dpscreenocr/
zlib License
231 stars 17 forks source link

Problems with oblique text #26

Closed ghost closed 9 months ago

ghost commented 1 year ago

Due to the nature of the program, it cannot detect oblique text or words, even if it detects it, it detects it incorrectly, an angled area selection can be beautiful.

danpla commented 1 year ago

Could you please attach an example image?

ghost commented 1 year ago

Normally I don't use English ocr a lot, but I realized this while trying the best data I just downloaded, you may have problems because you can't catch the angle in oblique texts, for example, here I tried to make the word bikinis to try ocr, it didn't work, I tried to take the sentence completely, it didn't work.

Screenshot_2023-01-22_05-27-33

danpla commented 1 year ago

The main problem with this image is not the rotation, but the font. Even if you fix the rotation manually, Tesseract will still not recognize the text, since the official English data was trained on more traditionally looking fonts such as Arial and Times. You will need to find third-party language data that was trained on fonts like the one in your image (or even do the training yourself, if you like).

ghost commented 1 year ago

@danpla yes I know, I threw this picture in a hurry, unfortunately I couldn't find a proper example, rotation change is still really important especially for those who want to ocr word by word, because you may have to ocr the whole sentence or paragraph because of the rotation problem. I know it can be hard to implement, but I believe it saves time in the long run.