hetwaterschapshuis/vivianite-segmentation

Vivianite segmentation

This model performs binary semantic segmentation of vivianite particles in RGB microscope images. It is distributed for use with the slib-analyser project, which provides a Gradio interface, batch inference, and particle measurements.

Intended use

Use the model to locate likely vivianite pixels, create binary masks, and support exploratory particle-area and particle-size analysis. Predictions should be reviewed by a domain expert before they inform operational or scientific decisions.

Model details

• Task: binary semantic segmentation
• Architecture: unet with a efficientnet-b0 encoder
• Input: 8-bit RGB microscope images
• Inference patch size: 128 × 128 pixels
• Default overlap: 50%
• Default probability threshold: 0.5
• Weights: SafeTensors

model_config.json contains the architecture metadata required to reconstruct the model. model.safetensors contains tensor weights only.

Usage

Install the project and launch the browser interface:

uv sync
uv run slib-analyser-ui

Run inference on a directory of PNG, JPEG, or TIFF images:

uv run python -m slib_analyser.training.inference \
  --image-dir path/to/images \
  --output-dir predictions

The first run downloads this release to the local Hugging Face cache. Later runs can reuse the cached files.

Evaluation

No evaluation metrics are included in this release. Validate the model on representative images before operational use.

Limitations

• The model is specialized for the image domain represented by its training data. Changes in microscope, illumination, magnification, sample preparation, or image processing can reduce accuracy.
• Nearby predicted regions are treated as one connected particle during measurement, while fragmented predictions are counted separately.
• Particle volume shown by the UI is not measured in 3D. It assumes each 2D connected component is a sphere with the same projected area.
• Threshold changes affect the resulting mask and all derived measurements.

Release files

• model.pth: model weights
• model_config.json: architecture and inference defaults
• README.md: this model card