HiReS

High-Resolution Image Segmentation and Analysis Pipeline

HiReS is a modular Python package and command-line tool for automated image segmentation and analysis.
It was designed for high-resolution biological or microscopy datasets, combining YOLO-based segmentation with geometry-aware postprocessing (via Shapely).

HiReS makes it easy to:

• Split large .tif or .png images into manageable Chunks
• Run instance segmentation on each Chunk
• Merge predictions seamlessly into global coordinates
• Filter and unify polygons
• Generate high-quality annotation overlays

---

Key Features

• Chunking: Divide large images into smaller overlapping Chunks for model inference
• Segmentation: Use YOLO-based instance segmentation on image chunks
• Filtering: Automatically remove edge-touching or invalid polygons
• Merging: Unify chunk-level detections into full-resolution space
• NMS: Apply IoU-based Non-Maximum Suppression on polygons
• Visualization: Create clear overlays showing detected regions
• Parallel Execution: Run multiple images simultaneously
• Cross-platform: Works on Linux, macOS, and Windows

---

Installation

# Clone and install
git clone https://github.com/StevetheGreek97/HiReS.git
cd HiReS
pip install -e .

or install directly from PyPI:

pip install hires

Requirements: Python ≥ 3.10
HiReS automatically installs Ultralytics, Shapely, Pillow, NumPy, and Matplotlib.
For GPU inference, install PyTorch separately from pytorch.org.

---

Command-Line Interface (CLI)

Once installed, HiReS provides a terminal command called hires.

General usage

hires <command> [options]

Available Commands

Command	Description
hires chunk	Split a large image into smaller Chunks
hires run	Run the full segmentation pipeline (single image or directory)
hires plot	Visualize segmentation overlays

---

hires chunk

Split an image into evenly sized Chunks for segmentation.

hires chunk --image raw_image.tif --out chunks/ --chunk-size 1024 1024 --overlap 150

Arguments:

Flag	Description	Default
--image	Path to a single image file	—
--out	Output directory for Chunks	—
--chunk-size	Chunk size in pixels: width height	1024 1024
--overlap	Overlap in pixels between Chunks	150

---

hires run

Run the complete segmentation pipeline on one image or a folder of images.
This automatically performs chunking, segmentation, filtering, merging, NMS, and visualization.

hires run --image data/ --model models/DaphnAI.pt --out results/ --workers 4

Pipeline steps:

1. Chunk input images
2. Predict segmentations using YOLO
3. Filter polygons touching image edges
4. Merge chunks into full-image coordinates
5. Apply IoU-based polygon NMS
6. Save final annotation file and visualization overlay

Arguments:

Flag	Description	Default
--image	Image file or directory of images	—
--model	Path to YOLO model (.pt)	—
--out	Output directory	—
--conf	Model confidence threshold	0.5
--imgsz	Inference image size	1024
--device	Compute device: cpu, cuda:0, or mps	cpu
--chunk-size	Chunk size (width height)	1024 1024
--overlap	Chunk overlap (pixels)	150
--edge-thr	Border-touch filtering threshold	1e-2
--iou-thr	IoU threshold for NMS	0.7
--workers	Number of parallel workers (for directories)	1

Outputs:

• results/<image>.txt → YOLO-style segmentation annotations
• results/<image>_annotated.tif → segmentation overlay image

---

hires plot

Overlay YOLO-format segmentation polygons on the original image.

hires plot --image raw_image.tif --ann results/raw_image.txt --out overlay.png

Arguments:

Flag	Description
--image	Path to the input image
--ann	YOLO-format annotation file
--out	Path to save the rendered overlay
--model	(Optional) model name/path for color consistency

---

Python API Example

If you prefer working from Python, HiReS can be used programmatically:

from HiReS.config import Settings
from HiReS.pipeline import Pipeline

cfg = Settings(
    conf=0.58,
    imgsz=1024,
    device="cpu",
    chunk_size=(1024, 1024),
    overlap=300,
    edge_threshold=0.01,
    iou_thresh=0.7
)

Pipeline(cfg).run(
    input_path="data/images/",
    model_path="models/DaphnAI.pt",
    output_dir="results/",
    workers=4
)

---

Project Structure

HiReS/
├── anno/              # Annotation parsing, filtering, NMS
├── ios/               # Chunking, plotting, writer, and YOLO inference
├── config.py          # Dataclass for pipeline settings
├── pipeline.py        # Main pipeline entry point
├── cli.py             # Command-line interface
└── tests/             # Example notebooks

---

Dependencies

HiReS depends on:

• ultralytics ≥ 8.0.0
• shapely ≥ 2.0.0
• Pillow ≥ 10.0.0
• numpy ≥ 1.25.0
• matplotlib ≥ 3.8.0

Optional (for GPU):

• torch (install via PyTorch.org)

---

Author

Stylianos Mavrianos
University of Hamburg
stylianosmavrianos@gmail.com

---

License

Licensed under the MIT License.
See the LICENSE file for details.

---