Thank you very much for your source code!
However, I met problem when I run your code again. I use datasets as you mentioned. convert to patchs 256x256 and augmentation rotation 90, 180 and 270(and I created 18792 new patchs for training) and train. My F-measure on Dibco2013 is 79,73% that is so far from paper(97.8%). So my question is:
Do you only use image from TA Gan or both fake image of TA Gan and original image for training?
Do you have using other technique?
Can you guess what I am wrong ?
Thank you!
Thank you very much for your source code! However, I met problem when I run your code again. I use datasets as you mentioned. convert to patchs 256x256 and augmentation rotation 90, 180 and 270(and I created 18792 new patchs for training) and train. My F-measure on Dibco2013 is 79,73% that is so far from paper(97.8%). So my question is: Do you only use image from TA Gan or both fake image of TA Gan and original image for training? Do you have using other technique? Can you guess what I am wrong ? Thank you!