RF-DETR Seg-Preview: Historical Document Instance Segmentation
This model is trained to detect and segment text lines and text regions from historical handwritten documents spanning from the 16th to the 20th century.
Model Description
RF-DETR Seg-Preview is an instance segmentation model based on the RF-DETR architecture. It is trained on Roboflow's rfdetr-library. More information about the architecture can be found via the link.
It predicts:
- Bounding boxes for text elements
- Class labels (text_region or text_line)
- Instance segmentation masks
Classes
The model detects two classes:
- text_region (index: 1) - Larger regions of text content
- text_line (index: 2) - Individual lines of text
Training Data
The model was trained on historical handwritten documents with the following data distribution:
- Training set: 11,495 images
- Validation set: 2,711 images
- Test set: 2,340 images
Performance Metrics
Validation Set Performance
| Class | mAP@50:95 | mAP@50 | Precision | Recall |
|---|---|---|---|---|
| text_region | 0.822 | 0.963 | 0.949 | 0.940 |
| text_line | 0.621 | 0.936 | 0.957 | 0.940 |
| Overall | 0.721 | 0.950 | 0.953 | 0.940 |
Test Set Performance
| Class | mAP@50:95 | mAP@50 | Precision | Recall |
|---|---|---|---|---|
| text_region | 0.822 | 0.959 | 0.949 | 0.940 |
| text_line | 0.688 | 0.955 | 0.978 | 0.940 |
| Overall | 0.755 | 0.957 | 0.964 | 0.940 |
Training Metrics
Use Cases
This model is particularly suitable for:
- Text line detection for OCR preprocessing
- Document digitization projects involving historical manuscripts
- Historical document understanding and analysis
Limitations
- The model is specifically trained on historical handwritten documents (16th-20th century)
- Performance may vary on modern printed documents or documents outside the training distribution
- Performance depends on image quality and document preservation state