Microplastics Detection

This repository contains a microplastics detection pipeline built around a residual U-Net-style segmentation model. It can train a model, run inference on microscopy images, and export visual detections as bounding boxes.

Highlights

• Segmentation-based microplastics detection.
• Training and inference scripts are included.
• A bundled model artifact is provided for immediate testing.
• Detection outputs both a rendered image and summary stats.

Repository Layout

• train_model.py: trains the segmentation model and writes the best checkpoint.
• det_p.py: loads a trained model and runs detection on a single image.
• standardize_masks.py: optional utility for normalizing mask values and polarity.
• myModel.keras: bundled model artifact.
• my_custom_resunet.keras: trained model checkpoint produced by training.
• detection_calibration.json: saved threshold calibration used by inference.

Requirements

Use Python 3.10+ and install the runtime dependencies:

pip install tensorflow numpy matplotlib opencv-python tifffile

If you are using a virtual environment, create and activate it first.

Quick Start

1. Clone the repository.
2. Install the dependencies.
3. Run training or inference from the project root.

Training

Train the model with:

python train_model.py

Training writes these outputs:

• my_custom_resunet.keras
• custom_resunet_training.png
• detection_calibration.json

If you want to standardize mask files before training, run:

python standardize_masks.py --masks-dir data/masks --output-dir data/masks_standardized

Inference

Run detection on a single image with:

python det_p.py --image path/to/image.png

Optional flags:

• --model path/to/model.keras: use a specific model file.
• --domain auto|clam|spiked: choose the inference profile.
• --no-calibration: ignore detection_calibration.json.
• --score-threshold 0.15: override the base score threshold.
• --color-threshold 0.05: override the base color threshold.
• --output detected_boxes.png: set the output image path.

Inference writes:

• detected_boxes.png

Notes

• The repository is set up for microplastics detection and visualization, not for a specific dataset walkthrough.
• If you train your own data, make sure image and mask pairs follow the filename stem convention used by the scripts.
• The detector will prefer the saved calibration file when available.

License

This project is released under the MIT License. See LICENSE for the full text.

Third-party library credits and license notices are listed in THIRD_PARTY_LICENSES.md.