SyntheticDID

This documentation will cover how to correctly generate synthesized images for training neural-networks

Requirements: OpenCV, Python(currently using 3.5.2), Java(currently using 1.8.0_101)

References: https://diuf.unifr.ch/main/hisdoc/divadid-document-image-degradation

1. script_generator.py *The python file takes in the following parameters: 1. number of outputs 2. stain level(float:1-5) 3. text noisiness level(float:1-5)

   *And performs the following actions:

    1.Generates the XML file, which then gets run by using the following command using
    DivaDid:
        java -jar DivaDid.jar data_generator_script.xml
    Running this script will generate synthesized degraded documents(in "data/outputs")
   
    2.Generates the blank images(in "data/blank_bgs/") with the sizes matching
    the randomly chosen background images, which will be used to generate 
    ground truths files(in "data/ground_truths")
   
    3. Generates another XML file named "match_generator_script.xml", which gets
    run by using the following command:
        java -jar DivaDid.jar match_generator_script.xml
    This command performs the same actions as the "data_generator_script.xml", 
    except it's on the corresponding blank backgrounds, right after this, you 
    should run:
        python ground_truth_binarization_script.py
    to finish generating the ground truth images which are in "data/ground_truths".
   

   *To specify the paths for the folders containing your word images, edit paths in "paths/word_image_folder_paths.txt", putting word folders in "data/input_words" is recommended.

   *To specify the paths for the folders containing your background images, edit paths in "paths/word_bg_folder_paths.txt", putting background folders in "data/input_bgs" is recommended.

   *To specify the paths for the folders containing your stain patches, edit paths in "paths/stain_folder_paths.txt", putting stain image folders in "data/stains" is recommended.

   *The python file also takes in the text noisiness level and apply a series of transformations to each randomly chosen word. The transformed images are located at "data/transformed_words".

2. clear.py *For convenience, this file clears all the images in "data/blank_bgs/", "data/ground_truths", "data/outputs" and "data/transformed_words". So user can regenerate images without having to hand-delete the images before.

3. ground_truth_binarization_script.py *This python file converts the DivaDid version of the ground truth to black & white. Since the raw images created by command "java -jar DivaDid.jar match_generator_script.xml" are not black & white. All the ground truth files are in "data/ground_truths"

Example: To generate 5 synthesized images with stain level = 1 and text noisiness level = 1, run:

    python script_generator.py 5 1 1
    
    java -jar DivaDid.jar data_generator_script.xml
    
To generate the corresponding ground-truth

    java -jar DivaDid.jar match_generator_script.xml
    
    python ground_truth_binarization_script.py
    
Repeat the process to generate more images with different parameters.