Ground-Truths

Git Link

350 actual screenshots from the various documents with accompanying transcriptions.

I measured the Zipf-Score and found it was high, with around 1.23 (Calculations done in survey_data.py). However, due to the rarity of some characters, it was difficult to capture it all. But I tried my best.

stu'ᵘshä'pmät. hɔi ɫä häpi'lumxʷ äku'stus xuic

Synthetic Data

1,000 'fake' data

Due to be a one-woman team, I hadn't the time to take a significant amount of ground-truth pairs. Because of that, I decided to create synthetic data using various Python scripts and packages found here

• Markovify: Using existing text, create novel words in the language using a Markhov chain model that captures patterns in existing words and recombines them Code
• Example words include: xʷä'ntc, q́ʷᴇńtsu,tcɩtśä'ᵃ̈
• Augraphy Allows you to alter images to add noise (blurriness, ink bleed, photocopy lines) Code
• Because Tesseract requires relatively clean images, I limited the amount of noise to just trying to replicate the typewriter like style
• Low Random Ink Lines: Adds ink lines randomly through the image
• Inkbleed: Captures all edges (ie letters) in the image and adds a slight blur
• Letterpress: Mimics uneven ink dispertion on the image
• Subtle Noise: Emulates the imperfections in scanning solid colors due to subtle lighting differences
• Pillow: Utilizing the Image functionality was able to create the images. Code

What is Tesseract?

Tesseract is an open-source OCR (Optical Character Recognition) engine using a Long Short-Term Memory neural network which is a special type of Recurrent Neural Network. It accepts an image (PNG, JPG) and outputs the image's text using various methods. Currently, Tesseract has 100+ languages, including some Indigenous languages.

What is Tesstrain?

The official training framework/toolkit for creating and fine-tuning Tesseract OCR models and allows you to train custom models for specific fonts, languages, or document styles not well-covered by default Tesseract models

To train it you have ground truth data (image + text pairs) as input to teach the model what characters look like in your specific context

Tesstrain allows fine-tuning from an existing pre-trained model (transfer learning) rather than training from scratch, which saves significant time and data

My Training

Did Multiple runs with different data:

• 350 ground-truths
• 350 gt + 500 synthetic pairs
• 350 gt + 1000 synthetic pairs

Paramaters

• 20,000 Iterations (number of times the model runs over the data)
• Default is 10,000 but it wasn't reaching ideal learning rate at 10,000
• Start Model: Latin script (this is a script that included a unicharset of all Latin characters which are captured in orthography)
• Doing without a base model and training from scratch would take 10,000 + gt