|
--- |
|
datasets: |
|
- Riksarkivet/trolldomskommissionen_seg |
|
- Riksarkivet/svea_hovratt_seg |
|
- Riksarkivet/krigshovrattens_dombocker_seg |
|
- Riksarkivet/jonkopings_radhusratts_och_magistrat_seg |
|
- Riksarkivet/gota_hovratt_seg |
|
- Riksarkivet/frihetstidens_utskottshandlingar_seg |
|
- Riksarkivet/bergskollegium_relationer_och_skrivelser_seg |
|
- Riksarkivet/bergskollegium_advokatfiskalskontoret_seg |
|
tags: |
|
- instance segmentation |
|
- handwritten |
|
- htr |
|
- text lines |
|
|
|
library_name: htrflow |
|
--- |
|
# Yolov9-lines-within-regions-handwritten |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
A yolov9 instance segmentation model for segmenting text-lines within text-regions. To be used after text-regions segmentation as part of an HTR-pipeline |
|
|
|
|
|
## Model Details |
|
|
|
### Model Description |
|
|
|
This model was developed for segmenting text-lines within text-regions in handwritten running-text documents. It wont work on full pages but only on cropped text-regions. |
|
It is meant to be implemented in an HTR-pipeline where one first segment text-regions, then text-lines within the regions, and then feed these text-lines to an HTR-model. |
|
|
|
- **Developed by:** The Swedish National Archives |
|
- **Model type:** yolov9 |
|
- **License:** [More Information Needed] |
|
|
|
### Model Sources |
|
|
|
- **Repository:** [yolov9-lines_within_regions-1](https://huggingface.co/Riksarkivet/yolov9-lines-within-regions-1) |
|
- **Paper:** [YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information](https://arxiv.org/abs/2402.13616) |
|
|
|
## Uses |
|
|
|
|
|
### Direct Use |
|
|
|
Segment text-lines within text-regions as part of an HTR-pipeline for handwritten running-text documents |
|
|
|
### Downstream Use |
|
|
|
As part of an HTR-pipeline for transcribing entire pages of handwritten running-text documents. See [Swedish Lion Libre](https://huggingface.co/Riksarkivet/trocr-base-handwritten-hist-swe-2) |
|
for example usage with the [HTRflow package](https://github.com/AI-Riksarkivet/htrflow) |
|
|
|
|
|
## How to Get Started with the Model |
|
|
|
### How to Load and Use the YOLOv9 Instance Segmentation Model |
|
|
|
Below is the Python code to load and use the trained YOLOv9 instance segmentation model using the Ultralytics repo: |
|
|
|
```python |
|
import torch |
|
from ultralytics import YOLO |
|
# Load the trained YOLOv9 model |
|
model = YOLO('path/to/your/model.pt') # Update with the correct path to your trained model |
|
# Load an image |
|
img = 'path/to/your/image.jpg' # Update with the path to the image you want to use |
|
# Perform instance segmentation |
|
results = model(img) |
|
# Display results |
|
results.show() # Show image with predicted masks |
|
# To get the raw predictions (bounding boxes, masks, etc.) |
|
for result in results: |
|
print(result.boxes) # Bounding boxes |
|
print(result.masks) # Segmentation masks |
|
``` |
|
|
|
### Usage with the HTRflow package |
|
|
|
See the model card for [Swedish Lion Libre](https://huggingface.co/Riksarkivet/trocr-base-handwritten-hist-swe-2) |
|
for example usage with the [HTRflow package](https://github.com/AI-Riksarkivet/htrflow), or refer to the documentation for |
|
[HTRflow](https://github.com/AI-Riksarkivet/htrflow) |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
[Trolldomskommissionen](https://huggingface.co/datasets/Riksarkivet/trolldomskommissionen_seg) |
|
[Svea Hovrätt](https://huggingface.co/datasets/Riksarkivet/svea_hovratt_seg) |
|
[Krigshovrättens domböcker](https://huggingface.co/datasets/Riksarkivet/krigshovrattens_dombocker_seg) |
|
[Jönköpings rådhusrätt och magistrat](https://huggingface.co/datasets/Riksarkivet/jonkopings_radhusratts_och_magistrat_seg) |
|
[Göta hovrätt](https://huggingface.co/datasets/Riksarkivet/gota_hovratt_seg) |
|
[Frihetstidens utskottshandlingar](https://huggingface.co/datasets/Riksarkivet/frihetstidens_utskottshandlingar_seg) |
|
[Bergskollegium relationer och skrivelser](https://huggingface.co/datasets/Riksarkivet/bergskollegium_relationer_och_skrivelser_seg) |
|
[Bergskollegium advokatfiskalkontoret](https://huggingface.co/datasets/Riksarkivet/bergskollegium_advokatfiskalskontoret_seg) |
|
|
|
|
|
### Training Procedure |
|
|
|
|
|
#### Training Hyperparameters |
|
|
|
See [training config](https://huggingface.co/Riksarkivet/yolov9-lines-within-regions-1/blob/main/args.yaml) at model repo |
|
|
|
|
|
## Evaluation |
|
|
|
See [training results](https://huggingface.co/Riksarkivet/yolov9-lines-within-regions-1/blob/main/results.csv) |
|
|
|
|
|
#### Metrics |
|
|
|
Standard metrics for instance segmentation. Note that evaluation of segmentation as part of an HTR-pipeline should be measured by what effect it |
|
has on the following HTR, that is, CER and WER for the entire page. For implementation and evaluation of entire HTR-pipelines, please check out [HTRflow](https://github.com/AI-Riksarkivet/htrflow), |
|
the Swedish National Archive's open-source package for HTR and OCR projects. |
|
|
|
|
|
### Model Architecture and Objective |
|
|
|
yolov9 |
|
|
|
#### Software |
|
|
|
[Ultralytics](https://github.com/ultralytics) |
|
|
|
## Citation |
|
|
|
[YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information](https://arxiv.org/abs/2402.13616) |
