ocr-lowresolution

Source code for low resolution OCR experiments

The source for the low-resolution version of Tesseract can be found at https://github.com/juliangilbey/tesseract/tree/lores-v1.0

More information can be found in our preprint here: https://arxiv.org/abs/2105.04515

Overview of the code

There are two separate tasks that the code in this repository performs:

• Creating Tesseract traineddata files for low resolution images
• Running Tesseract on low resolution images to perform OCR

We describe each of these in turn; you first need to build and install the modified version of tesseract pointed to above to use the code here.

There is also a file called scripts.sh which contains a variety of (briefly commented) shell commands that can be adapted to run the code to perform the training and OCR.

Creating Tesseract traineddata files

As described in the preprint, pre-trained traineddata files are available on Zenodo, so this is only required if you are reproducing the experiments.

Creating traineddata files is achieved using make. You will need to modify the start of the Makefile for your local setup. In particular:

• Modify the TESSTRAIN line at the start of the Makefile to point to a directory in which you will store all of your training data.

• You may wish to modify the TESSDATA if you wish to store the training data elsewhere; by default it is $(TESSTRAIN)/tessdata.

• You will have to change the TESSCONFIGS line to point to the location of tesseract config files. The version in GitHub works for macOS with MacPorts, while for a Linux distribution, /usr/share/tesseract-ocr/4.00/tessdata/configs may be more appropriate (assuming that the locally built modified Tesseract is installed overwriting the standard version).

• If your low-resolution tesseract executable is not in your PATH (before any non-modified tesseract), you will need to set TESSBINDIR to the directory containing the low-resolution tesseract.

• Depending on the system you are using, the available fonts may be different. Modify the FONTS setting if so. These fonts must work with LuaLaTeX.

Before running make for the first time, you will need training text data split into manageable chunks. To replicate the experiments, download Tesseract's own training data from https://github.com/tesseract-ocr/langdata_lstm/raw/main/eng/eng.training_text and save it in the directory $(TESSDATA)/eng. Then in the directory $(TESSDATA)/eng, run the command: split -a 3 -l 160 eng.training_text eng.training_text_split_

(If you want to use different training texts, replace eng.training_text in this command with a different file of training text.)

The training has been split into three steps, which must be run in order:

• make training-images: this makes low-resolution images from the training text
• make lists: this creates some intermediate files necessary for the training step
• make training: this performs the actual training

By default, the resolution is 60 dpi, the scaling method is bicubic and there is no Gaussian blur. (See the preprint for more details on these.) These can be changed by setting Makefile parameters on the command line as follows:

make RES=75 SCALING=0 BLUR=1 training-images

(and the same parameters should be used for make lists and make training). The RES parameter specifies the resolution, SCALING is 0 for box, 1 for bilinear and 2 for bicubic, and BLUR specifies the Gaussian blur in pixels.

The directory which will contain the resulting traineddata file is $(TESSTRAIN)/data$(RES)_$(SCALING_NAME)+$(BLUR)

It is also possible to run the training in steps using the trainingstep target; see the Makefile for more details. There are also some clean targets; again, see the Makefile.

Performing OCR on low-resolution images

This is achieved using the script ocr_images.py. Running ocr_images.py --help describes all of the commandline options. Here is an example of how it can be used:

ocr_images.py -r 60 -s C0,L0,B0,B0.5,B1,B1.5,B2 -g image.gt.txt \
    --outbase image --tessdata-path /path/to/tessdata/ -w image.png

This processes the 60 dpi image image.png, whose known ground truth text is image.gt.txt. The low-resolution traineddata files are in the directory /path/to/tessdata; this should match the value of TESSDATA used in the Makefile in the first stage. The output files will have the prefix image, and -w specifies that word-level metrics will be generated, specifying how accurate the OCR performance was. Finally, this command will process the image multiple times and merge the results: it will use the C0 network (bicubic with no further Gaussian blurring), then L0 (bilinear with no further Gaussian blurring), then B0, B0.5, ..., B2 (box with Gaussian blurring of 0, 0.5, ..., 2 pixels standard deviation). This requires the presence of traineddata files for each of these scalings.