ImageOperations is a minimal, device‑aware image ops library built on top of OpenCV and NumPy. You can call functions directly without initialization, or use a small facade class with a default device (CPU/GPU/MPS placeholder). When the requested device is unavailable, the library transparently falls back to CPU.

• ✅ Functional API and OOP API
• ✅ Device selection per call (device="cpu"|"gpu"|"mps")
• ✅ Safe fallback (GPU/MPS → CPU)
• ✅ Typed (PEP 561, py.typed)
• ✅ Clean, modular structure (ops/, registry, device, types)
• 🔧 Easy to extend: add new ops or backends without touching the public API

Note: MPS (Apple Metal) is reserved in ComputeDevice but currently falls back to CPU.

poetry add img-ops
# or with pip:
# pip install img-ops

Headless environments: replace opencv-python with opencv-python-headless in your environment if you don't need GUI capabilities.

from img_ops import bgr_to_hsv, in_range, gauss_blur, ComputeDevice

hsv = bgr_to_hsv(img, device="gpu")  # uses GPU if CUDA available; else CPU
mask = in_range(hsv, (0, 80, 80), (10, 255, 255))  # default device (CPU)
blur = gauss_blur(img, (5, 5), 1.2, device=ComputeDevice.CPU)

from img_ops import ImageProcessor

proc = ImageProcessor(default_device="gpu")
hsv = proc.bgr_to_hsv(img)  # defaults to GPU (fallback → CPU)
mask = proc.in_range(hsv, (0, 80, 80), (10, 255, 255))  # uses proc.default_device
blur = proc.gauss_blur(img, (5, 5), 1.2, device="cpu")  # per-call override → CPU

• Color: bgr_to_rgb, bgr_to_hsv, in_range
• Filters: gauss_blur
• Segmentation (CPU-only): find_contours, connected_components
• Utils: sum_masks (internal registry op)

Add more ops by editing registry.py and adding thin wrappers in ops/.

from img_ops import ComputeDevice
# ComputeDevice.GPU  -> CUDA (if available)
# ComputeDevice.MPS  -> reserved (falls back to CPU)
# ComputeDevice.CPU  -> always available

• If you request GPU and CUDA is unavailable, the call automatically falls back to CPU.
• MPS is a placeholder and currently always falls back to CPU.

The library ships with type hints and py.typed.

from img_ops import CvImage, PointHSV

# CvImage: NDArray[uint8 | float32]  # shape (H,W,3) or (H,W)
# PointHSV: tuple[int, int, int]     # (H, S, V)

Shape is not fixed in the type system; keep (H, W) or (H, W, 3) by convention.

• Unknown device strings raise ValueError with allowed values.
• Missing implementations raise NotImplementedError (after attempting CPU fallback).

1. Implement CPU/GPU/MPS callables in registry.py under _cpu(), _gpu(), _mps().
2. Add a thin function wrapper in ops/<category>.py that normalizes device and calls dispatch(...).
3. (Optional) Expose it in ops/__init__.py and the package __init__.py.

1. Extend ComputeDevice (e.g., OPENCL, TORCH).
2. Add a registry builder (e.g., _opencl()), wire it in REGISTRY.
3. Add availability checks in device.py and normalization in each ops/*.py wrapper.

This keeps public API stable while evolving internals.

• GPU helps mostly for larger images or heavy kernels; small images can be faster on CPU due to upload/download overhead.
• Prefer cv2.add over a + b for masks to handle saturation properly.

poetry run pytest
poetry run mypy src/img_ops
poetry run ruff check .

Project: https://github.com/omigutin/img_ops

Project Tracker: https://github.com/users/omigutin/projects/4

Contact: migutin83@yandex.ru

MIT License. See LICENSE for details.