About: Bleed-Through Removal from Degraded Documents

[kerberosargos]

Firstly thank you for your great work. I am newbie about AI. I need your help.

My question is about "Bleed-Through Removal from Degraded Documents" I have some images for testing. Your Docres "appearance" task almost works perfectly. I would like to clean images than try to detect text with "Craft" model. But some transparent object in image, causes problems. Some samples are as following. Can you give a advice for removing transparent objects for good text detection?

Thank you in advance for your effort.

Orginal	DocRes	Craft

Orginal	DocRes	Craft

Orginal	DocRes	Craft

[ZZZHANG-jx]

Firstly thank you for your great work. I am newbie about AI. I need your help.

My question is about "Bleed-Through Removal from Degraded Documents" I have some images for testing. Your Docres "appearance" task almost works perfectly. I would like to clean images than try to detect text with "Craft" model. But some transparent object in image, causes problems. Some samples are as following. Can you give a advice for removing transparent objects for good text detection?

Thank you in advance for your effort.

Thank you for your interest in our work. For the bleed-through degradation, we primarily simulated it in our training data using the bleed_through function. In the images you provided, the bleed-through is quite severe. I believe the simulation in the training data needs to be more intense, such as reducing the value in this line from 0.75 to a smaller number, and then fine-tuning the model. Additionally, you might want to try our other work, GCDRNet, which is specifically designed for appearance enhancement and might yield better results.

[kerberosargos]

Firstly thank you very much for your quick answer. Acctually I have tried GCDRNet before DocRes with same test images. But DocRes result is better than GCRNet for me.

As I understand from your answer, I should train again your used dataset by changing 0.75 value. am I right?

[ZZZHANG-jx]

Firstly thank you very much for your quick answer. Acctually I have tried GCDRNet before DocRes with same test images. But DocRes result is better than GCRNet for me.

As I understand from your answer, I should train again your used dataset by changing 0.75 value. am I right?

Yes, that's correct. If you're not concerned with other tasks (such as binarization or dewarping), you can train only the appearance task, which will simplify things.