nzIng/unet-manuscripts-segmentation

U-Net Manuscript Line Segmentation

TensorFlow/Keras U-Net model for line segmentation on medieval manuscript page images.

Files

• unet-segmentation.keras: trained U-Net checkpoint.
• inference.py: simple local inference script.

Usage

python inference.py --input-path path/to/page.jpg

Outputs are created automatically:

• results_json_inference/inference_<image_name>.json
• crops_line/inference_<image_name>/line_XXXX.png

To process a folder:

python inference.py --input-path path/to/images_dir

Output

The JSON contains page metadata and detected text lines with:

• bbox
• polygon
• score
• reading_order
• optional crop metadata

Notes

The model input size is 256 x 256 x 3. The predicted mask is post-processed into line bounding boxes and line crops.