cqtan / ml-cnn-unwanted

Filters out unwanted Stabi images and outputs them in separate directories, here in the 'output2/content' directory

CNN model for filtering out unwanted Stabi images

Created specifically to group Stabi images together that have meaningful content. This way less data is needed to be saved in the databank and clustering may be performed only on images with actual content. Furthermore, the TIFF images are converted to JPEG.

---

Dependencies:

Dependencies used for this are listed in the requirements.txt file. Download them easily to your current Python environment (or switch to virtual one) by navigating to the project root and entering the following:

• pip install -r requirements.txt

Or when using Anaconda:

• First, add these to the conda channels to install packages outside from the default channels:
  • conda config --append channels conda-forge
  • conda config --append channels gilbertfrancois
• Then install the packages listed in the 'package-list.txt' with:
  • conda install --file package-list.txt

---

tl;dr:

Run the following to first download SSB images:

• python sbbget.py

Once all images from the OCR-PPN-Liste.txt have been downloaded to sbb/saved_images, run:

• python classify-all.py

Output files output1 and output2 will be created but only the latter is of concern. Meaningful content should then be in output2/content. Double check other directories for outliers though!

If any of the scripts are interrupted, simply rerun the interrupted script. It should pick up where it left off or when in doubt, check the last section about Output and Logging.

---

Preparing data:

The following sections explain in more detail the workflow of the scripts.

The script has been configured to locate the images by path, here being the ssb directory. Simply place all subdirectories within this directory. The subdirectories should then contain the images to be filtered. An example directory structure is as follows:

• sbb
  • saved_images
    • PPN61019657X
      • 0001_Page1_Block2.tif
      • 0001_Page1_Block12.tif
      • 0132_Page1_Block9.tif
      • ...
    • PPN610195530
      • ...
    • PPN610195867
      • ...
    • PPN610196898
      • ...
    • ...

---

Running the script:

The only script you have to run is the 'classify-all' script. The arguments to be passed are as follows:

• model: path to trained model model.
• labelbin: path to label binarizer.
• images: path to input images.

The current version should run with the following command if images are properly placed in the 'sbb' directory as suggested above:

python classify-all.py

Once the script is done, 'output' directories will be created with subdirectories according to the number of classes the model was trained with.

Due to imbalance in training data, two models were trained to filter the images. The first model filters the following:

• blanks (Blank pages)
  • ca. 1500 training data.
• color_palette (An image of a color palette)
  • ca. 200 training data.
• content (The images the we want)
  • ca. 1700 training data.

The second model filters the following:

• bar_code (Stabi bar codes)
• covers (Simple book covers, complex ones are left out)
• logo (The Stabi logo)
• red_stamp (Red Stabi stamp mark)
• content (The images the we want)

Note:

A single image may also be classifed for debugging purposes by running the 'classify-single' script. Arguments for this are as follows:

python classify-single.py --model filter_v2.model --labelbin label_v2.pickle --image data/0492_Page1_Block1.tif

Make sure the last argument '--image' is pointing towards a single image file.

---

Output and Logging:

The output1 and output2 directories will be created (if not already). Images will be grouped by their predicted class followed by their directory name (usually the PPN name). The file names themselves remain the same as well though they will instead be converted to JPEG iamges to save space.

Log files will also be created to keep track of the progress of both scripts. Make sure to check them if errors occur or when scripts are interrupted.

The first script will continue where it left off if unterrupted and if the respective ppn_log.txt has been created on the first run.

The second script also continues where it left off if interrupted by reading both the output1_log.txt and output2_log.txt. These two will simply list the PPNs that it has already dealt with and check against the newly read. Simply rerun python classify-all.py if this happens.

The LOG_2018_.log simply lists more information on the filtering progress for each run of the classify-all.py script.