search

ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
766 stars 254 forks source link

Can pixel link only detect single word, not long line? #96

Open jisheng047 opened 5 years ago

austingg commented 5 years ago

It's hard to say what is a word only by image content. Because, the space between character is relative.

jisheng047 commented 5 years ago

@austingg using like ctpn can detect correct about long-lines. But i am working on scene text recognition project, include 2 phase (detect + recognition). The second phase only work with word, poor performance for long-lines. I have tried to use Seg-Link, EAST but the result is not good. @@! I just want to word partition for text. Do you have any idea to solve this problem ?

austingg commented 5 years ago

you may try to use higher pixel_score threshold or higher link_conf_threshold, which may be helpfuly. Moreover, you can try sequence recognition. @jisheng047

jisheng047 commented 5 years ago

@austingg I have tried it by increase the link_conf_threshold. The result now is better than before. I have tried sequence recognition by using ctpn (detection) and the attention ocr (recognition) previous, but the result was not good. Anyway thanks for your idea!

nassimaNoufail commented 5 years ago

@jisheng047 how did u do the recognition, I was able to get good results with this paper for text detection in images with long lines but I can't find the recognition part. do you have any idea how to do it?

jisheng047 commented 5 years ago

@nassimaNoufail you can use crnn or attention for recognition. Can you modify it for single word detection ?

Dgs29 commented 5 years ago

@jisheng047 @austingg where inside the code should I increase pixel_score threshold / link_conf_threshold and to what value or above should I set it?

austingg commented 5 years ago

@DGS There are threshold argument in the test_pixel_link.py file

Dgs29 commented 5 years ago

@austingg .. I 've increased it to pixel_conf_threshold & link_conf_threshold to 0.9. and the result is same, it only detects word by word . If I increase it past 1.0 it doesn't seem to work..