Game Image OCR Training Workflow - spectrumbranch/retro-translation-project GitHub Wiki

1. Convert source video files to images
2. Prepare images for OCR training
3. Prepare ground truth (training/verification data) for OCR.
4. Train OCR
5. Evaluate results.

Convert source video files to images

Convert source video files to still images.

• Potentially use detector to find rectangles in video, classifier to identify and generate only useful images (those with dialogue).

Prepare images for OCR training

Process images, generating "optimal" inputs for OCR engine.

Prepare ground truth (training/verification data) for OCR.

Find text areas within larger fullscreen game images. Generate inputs required by OCR engine

• For Tesseract, single line images and corresponding reference text.

Train OCR

Use generated ground truth as input for training session; train OCR model.

Evaluate results

Evaluate results of OCR training session against any previous training sessions. Use the results to improve variables for future training sessions.