predict results on real documents is bad

[huyutao3550346]

hello author, I have implemented the HTSNet by keras from the paper.
after train with 90 epochs and 30 steps IOU is nearby 95%
and model predict well in synthesized test set. like:
origin image:

result image:

but predict results on real documents is bad. like:
origin image:

result image:

here is my model define and train file:
model define:
https://drive.google.com/file/d/1XvIQswjj4iDwEowWw51eS1lOEPtjee_n/view?usp=sharing
predict.py:
https://drive.google.com/file/d/1aeLbIippxarnMvjEIjGie-0cx6n0SvP1/view?usp=sharing
train file:
https://drive.google.com/file/d/1V9IhUozXp_dEIO8boV9VNwbUbMqou3ea/view?usp=sharing

Is my loss function and model defined wrong or do I need to use real documents for training?
Do you have any good suggestions to solve this problem? thank you.