search

xiaoyu258 / DocProj

Document Rectification and Illumination Correction using a Patch-based CNN
MIT License
345 stars 87 forks source link

The dataset doesn't have the original scanned image #4

Closed kisstherainfh closed 4 years ago

kisstherainfh commented 5 years ago

Hi, when I follow your work, I find that the dataset you provide donesn't incorporate the scanned images, which are used as gorund truth to train the illumination correction network. Could you release this datavset

Thank you！

xiaoyu258 commented 4 years ago

Hi @kisstherainfh, sorry for my late reply. Please check the link for the scanned images: https://drive.google.com/open?id=1PKAtO3ml0FCXZ92j77SHtxo0XWX9QkVn

kunlaotou commented 4 years ago

Hi , I am lucky to find this paper and I want to train the illumination correction network to address blurred images problem in my OCR. How can I get distorted training images if I get the scanned images？using resampling.rectification function？？ by the way @kisstherainfh， How does the network you train perform in real scenarios? thx！！！

fh2019ustc commented 4 years ago

@xiaoyu258 @kunlaotou Hi, I have not trained the illumination correction network . The question you asked is also my question. When I get a distorted crop image, how can I get the corresponding ground truth in the scanned images to train the illumination correction network. The model I trained performed well in real scenarios. I sorry for my late reply！@kunlaotou

xiaoyu258 commented 4 years ago

For the correction network, you can use resampling.rectification function to correct the distorted image by the flow to get the input. For the ground truth scanned image, please find the link here: https://drive.google.com/open?id=1PKAtO3ml0FCXZ92j77SHtxo0XWX9QkVn