vietocr-animesub-finetune

VietSubOCR

<div align="center">

</div>

Model Description

A Vietnamese OCR model fine-tuned from VietOCR using the vgg_transformer configuration. The model uses VGG as the feature extractor and a Transformer for sequence recognition.

This model is trained on a custom dataset to recognize Vietnamese text from images, with a particular focus on anime subtitles.

Training

• Base model: VietOCR vgg_transformer
• Framework: PyTorch / VietOCR
• Training iterations: 10,000
• Train / Validation split: 90% / 10%
• Validation frequency: Every 1,000 iterations
• Validation samples: 1,000 samples

Evaluation

Full Sequence Accuracy measures the percentage of samples for which the entire predicted text sequence is correct.

Per Character Accuracy measures the accuracy at the individual character level.

Intended Use

This model is intended for recognizing Vietnamese text from images, particularly Vietnamese subtitles extracted from movies and videos.

Limitations

Performance may decrease when the input images contain fonts, sizes, colors, backgrounds, or layouts that differ significantly from the training data.

The gap between Per Character Accuracy (89.9%) and Full Sequence Accuracy (73.6%) indicates that the model can correctly recognize most individual characters while still making one or more character-level errors in some complete sequences.