HU-PageScan: a fully convolutional neural network for document page crop

This is an implementation of HU-PageScan on Python 3, Keras and Tensorflow. The training dataset is the Extended Smartdoc Dataset. The proposed model is based on U-net Architecture, which we halved the number of parameters.

Edge detection Train:

• Architecture used: HU-PageScan
• Input Image Dimensions: 512x512 (gray - 1 channel)
• Dimensions Output Image: 512x512 (binary - 1 channel)
• Pre-trained model provided: pre_trained_model.hdf5

Instructions for performing the training:

Enter the command below in cmd

python "run_hupagescan.py" -munet --train-folder=train_folder/" 
       --validation-folder="validation_folder" --gpu="1" --bs="4" --train-steps="1" 
       --valid-steps="1" --no-aug --train-samples="18598" --valid-samples="6165" --lr="0.0001"

• Set Amount of epochs;
• Set a location where new .hdf5 (trained model) will be saved
• Define the location of results during training (output_refined)

Instructions for converting .hdf5 to .pb:

Run the script: generator_hdf5_pb_pbtxt.py

Set:

• hdf5 path
• Destination path of .pb and .pbtxt files

Instructions for performing .pb file inference

Run the script: inferenceTestPB.py Define:

• The location where the output will be saved
• Location of the images to be inferred
• Locating the .pb
• Name of .pb

Extra information

• Input layer name: input_1:0
• Output layer name: conv2d_19/Sigmoid

Citation

Use this BibTeX to cite this repository:

@ARTICLE{
   author = {Ricardo Batista das Neves Junior},
   author = {Estanislau Lima},
   author = {Byron L.D. Bezerra},
   author = {Cleber Zanchettin},
   author = {Alejandro H. Toselli},
   ISSN = {1751-9659},
   language = {English},
   title = {HU-PageScan: a fully convolutional neural network for document page crop},
   journal = {IET Image Processing},
   year = {2020},
   month = {December},
   publisher ={Institution of Engineering and Technology},
   copyright = {© The Institution of Engineering and Technology},
   url = {https://digital-library.theiet.org/content/journals/10.1049/iet-ipr.2020.0532}
}