Albucore: High-Performance Image Processing Functions

Albucore is a library of optimized atomic functions designed for efficient image processing. These functions serve as the foundation for AlbumentationsX, an image augmentation library.

Overview

Image processing operations can be implemented in several ways, with performance depending on dtype, size, layout, and channel count. Albucore routes each operation to a benchmark-selected NumPy, OpenCV, NumKong, StringZilla, or eligible CPU PyTorch implementation.

Most image-processing routers support uint8 and float32. The elementwise exp, log, and sqrt routers are float32-only; conversion helpers also support additional integer dtypes.

Key features:

• Optimized atomic image processing functions
• Automatic selection of the fastest implementation based on input image characteristics
• Seamless integration with AlbumentationsX
• Reproducible micro-benchmarks and committed routing reports (see benchmarks/README.md)

Installation

Requires Python 3.10+. Choose and install the PyTorch build for your CPU, CUDA, or MPS environment first. Then install Albucore with an OpenCV extra. For a Linux CPU-only headless application:

pip install "torch>=2.13.0" --index-url https://download.pytorch.org/whl/cpu
pip install "albucore[headless,torch]"

CUDA or macOS (MPS): Select and install the matching PyTorch build with the PyTorch installation selector, then install Albucore:

pip install "albucore[headless,torch]"

pip install albucore installs only the base dependency set. Use it when a transitive consumer only resolves Albucore, such as during a documentation build. The current public import graph requires both OpenCV and PyTorch.

The torch extra declares Albucore's PyTorch runtime requirement but cannot select a CPU, CUDA, or MPS wheel through standard package metadata. Use PyTorch's platform-specific installation command before installing Albucore. The examples use the headless OpenCV extra; replace it with gui, contrib, or contrib-headless if needed.

AlbumentationsX passes prevalidated CPU, strided Torch tensors with requires_grad=False to resize3d and warp_affine3d; the low-level routers do not repeat those checks or move/detach Tensor data.

With OpenCV GUI support (for local development with cv2.imshow):

pip install "albucore[gui,torch]"

With OpenCV contrib modules:

pip install "albucore[contrib,torch]"              # GUI version
pip install "albucore[contrib-headless,torch]"     # Headless version

Note: If you already have opencv-python or opencv-contrib-python installed, use pip install "albucore[torch]" after installing the platform-specific PyTorch build. This does not add another OpenCV package; Albucore uses the existing installation.

Usage

import numpy as np
import albucore

# Create a sample RGB image
image = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)

# Apply a function
result = albucore.multiply(image, 1.5)

# For grayscale images, ensure the channel dimension is present
gray_image = np.random.randint(0, 256, (100, 100, 1), dtype=np.uint8)
gray_result = albucore.multiply(gray_image, 1.5)

Albucore automatically selects the most efficient implementation based on the input image type and characteristics.

Shape Conventions

Albucore expects images to follow specific shape conventions, with the channel dimension always present:

• Single image: (H, W, C) - Height, Width, Channels
• Grayscale image: (H, W, 1) - Height, Width, 1 channel
• Batch of images: (N, H, W, C) - Number of images, Height, Width, Channels
• 3D volume: (D, H, W, C) - Depth, Height, Width, Channels

Important Notes:

1. Channel dimension is always required, even for grayscale images (use shape (H, W, 1))
2. Single-channel images should have shape (H, W, 1) not (H, W)
3. Batch vs volume: (N, H, W, C) is N separate images; a single 3D volume is (D, H, W, C) with depth D. Do not confuse N (batch) with D (slices).

Examples:

import numpy as np
import albucore

# Grayscale image - MUST have explicit channel dimension
gray_image = np.random.randint(0, 256, (100, 100, 1), dtype=np.uint8)

# RGB image
rgb_image = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)

# Batch of 10 grayscale images
batch_gray = np.random.randint(0, 256, (10, 100, 100, 1), dtype=np.uint8)

# 3D volume with 20 slices
volume = np.random.randint(0, 256, (20, 100, 100, 1), dtype=np.uint8)

Functions

The tables below highlight commonly used public routers. They are exported via from albucore import *. The compatibility shims in albucore.functions cover only the names documented in docs/public-api.md; warp_affine3d is intentionally public from albucore and albucore.geometric only.

Image routers use channel-last inputs with an explicit channel dimension ((H, W, C), never bare (H, W)) and generally support uint8 and float32. Exceptions are stated in the tables.

Arithmetic

value / factor / exponent can be a scalar, a length-C 1-D array (per-channel), or a full image-shaped array.

These arithmetic routers do not impose an image-range convention on float32 data. The explicit @clipped decorator remains available for callers, including AlbumentationsX operations whose own contract requires clipping.

Elementwise math

These functions accept float32 arrays up to rank 4 and preserve the exact input shape. With inplace=True, an owned writable buffer may be reused; views and read-only arrays are never mutated. See the elementwise benchmark report for routing thresholds, environment, and NumKong results.

Normalization

normalize is for fixed per-channel constants (ImageNet-style). normalize_per_image estimates stats from the image at call time.

Statistics

axis accepts None/"global" (scalar), "per_channel" (shape (C,)), or any NumPy-style int/tuple[int, ...].

LUT (lookup tables)

Geometric / spatial

The package also star-exports multi-channel wrappers for copy_make_border, gaussian_blur3d, remap, resize, resize3d, separable_filter3d, warp_affine, warp_affine3d, and warp_perspective; see docs/public-api.md and their docstrings for complete signatures. gaussian_blur3d, separable_filter3d, resize3d, and warp_affine3d expect exactly one prevalidated NumPy DHWC volume or Torch CDHW tensor per call.

Type conversion

Decorators (re-exported)

See docs/decorators.md for @preserve_channel_dim, @contiguous, @clipped, and @batch_transform (used internally, not re-exported).

Array layouts and batch processing

Arithmetic, normalization, statistics, conversion, and elementwise routers operate on channel-last arrays and preserve these layouts where applicable:

• Single images: (H, W, C)
• Batches: (N, H, W, C)
• Single volumes: (D, H, W, C)

Spatial routers document their own image-shape requirements. Transform authors can use @batch_transform to adapt an image operation to documented array ranks while restoring the original layout.

See docs/decorators.md for internal decorator documentation (@preserve_channel_dim, @contiguous, @clipped, @batch_transform).

Performance

Albucore uses a combination of techniques to achieve high performance:

1. Multiple Implementations: Each function may have several implementations using NumPy, OpenCV, NumKong, StringZilla, or CPU PyTorch when the documented contract permits it.
2. Automatic Selection: The library chooses a backend from dtype, size, memory layout, channel count, and semantic constraints.
3. Measured Routing: Backend choices and thresholds come from repeatable benchmarks rather than backend preference.
4. NumKong: SIMD blend for uint8 and single-channel float32 add_weighted, plus same-shaped uint8 add_array; cdist for small pairwise_distances_squared; wide-accumulator moments for selected statistics routes (see docs/numkong-performance.md).
5. CPU PyTorch: Selected eager 3D volume operations use full-path Torch routes after benchmarks include the NumPy/Tensor bridge, thread settings, and output repair (see docs/torch-performance-optimization.md).

Micro-benchmarks vs NumPy/OpenCV/NumKong: see benchmarks/README.md. Run uv run python benchmarks/benchmark_elementwise.py for exp/log/sqrt, or uv run python benchmarks/benchmark_numkong.py for a smaller NumKong sweep.

See docs/performance-optimization.md for the general workflow and docs/torch-performance-optimization.md for eager CPU Torch routes.

Documentation

• CONTRIBUTING.md - Pull request process and CLA acceptance paths
• AGENTS.md - AI development guidelines for working with this codebase
• docs/image-conventions.md - Image shape conventions and requirements
• docs/decorators.md - Decorator usage and patterns
• docs/performance-optimization.md - Performance optimization guidelines
• docs/torch-performance-optimization.md - Eager CPU PyTorch routing and benchmark guidance
• docs/torch-tensor-migration-plan.md - Future CPU Tensor contract and integration gates
• docs/numkong-performance.md - Current NumKong routes and benchmark decisions
• docs/public-api.md - Star-exported routers vs albucore.functions shims
• benchmarks/README.md - Python micro-benchmarks (uv run python benchmarks/…)

License

Albucore is publicly available under the MIT License, including contributions accepted under the Albucore Contributor License Agreement Version 1.0. The CLA does not change the repository's public license. Historical contributions remain available under MIT and become CLA-covered only through an applicable Version 1.0 Acceptance Record. See CONTRIBUTING.md for the individual CLA Assistant and entity acceptance paths.

Acknowledgements

Albucore provides core image-processing primitives for AlbumentationsX. We'd like to thank all AlbumentationsX contributors and the broader computer vision community for their inspiration and support.