albucore 0.1.5

High-performance image processing functions for deep learning and computer vision.

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, a popular image augmentation library.

Overview

Image processing operations can be implemented in various ways, each with its own performance characteristics depending on the image type, size, and number of channels. Albucore aims to provide the fastest implementation for each operation by leveraging different backends such as NumPy, OpenCV, and custom optimized code.

Supported dtypes: uint8 and float32 only.

Key features:

• Optimized atomic image processing functions
• Automatic selection of the fastest implementation based on input image characteristics
• Seamless integration with AlbumentationsX
• Benchmark scripts available (see ./benchmark.sh <data_dir>)

Installation

Requires Python 3.10+. Basic installation (you manage OpenCV separately):

pip install albucore

With OpenCV headless (recommended for servers/CI):

pip install albucore[headless]

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

pip install albucore[gui]

With OpenCV contrib modules:

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

Note: If you already have opencv-python or opencv-contrib-python installed, just use pip install albucore to avoid package conflicts. Albucore will detect and use your existing OpenCV 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
• Batch of volumes: (N, D, H, W, C) - Number of volumes, 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)

# Batch of 5 RGB volumes, each with 20 slices
batch_volumes = np.random.randint(0, 256, (5, 20, 100, 100, 3), dtype=np.uint8)

Functions

All symbols below are exported via from albucore import * (or from albucore.functions import …). Every function accepts uint8 and float32 images; inputs must have an explicit channel dimension ((H, W, C) — never bare (H, W)).

Arithmetic

Function	Signature	What it does	How it works
multiply	(img, value, inplace=False)	img * value, clipped to dtype range	uint8 scalar/vector → LUT; uint8 array → OpenCV; float32 → NumPy broadcast
add	(img, value, inplace=False)	img + value, clipped to dtype range	uint8 scalar → OpenCV saturate; uint8 vector → LUT; uint8 array → NumKong/OpenCV; float32 → NumPy
power	(img, exponent, inplace=False)	img ** exponent, clipped to dtype range	uint8 → LUT; float32 scalar → cv2.pow; float32 array → NumPy
add_weighted	(img1, weight1, img2, weight2)	img1*w1 + img2*w2, clipped	NumKong SIMD blend; large float32 → cv2.addWeighted
multiply_add	(img, factor, value, inplace=False)	img * factor + value, clipped	uint8 → LUT (fused, one table); float32 → NumPy broadcast

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

Normalization

Function	Signature	What it does	How it works
normalize	(img, mean, denominator)	(img - mean) * denominator → float32	uint8 → LUT (256-entry float32 table per channel); float32 → NumPy fused. Caller-supplied constants (e.g. ImageNet stats).
normalize_per_image	(img, normalization)	Normalize using stats computed from img → float32	uint8 → LUT (except "min_max" → cv2.normalize); float32 → OpenCV/NumPy. normalization ∈ {"image", "image_per_channel", "min_max", "min_max_per_channel"}

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

Statistics

Function	Signature	What it does	How it works
mean	(arr, axis=None, *, keepdims=False, dtype=None)	Population mean	uint8 global → NumKong moments; per-channel HWC ≤ 4ch → cv2.mean; else NumPy
std	(arr, axis=None, *, keepdims=False, eps=1e-4, dtype=None)	Population std + eps	uint8 global → NumKong moments; per-channel HWC ≤ 4ch → cv2.meanStdDev; else NumPy
mean_std	(arr, axis=None, *, keepdims=False, eps=1e-4)	Mean and std+eps jointly	Single NumKong moments pass for uint8 global (faster than separate mean+std)
reduce_sum	(arr, axis=None, *, keepdims=False)	Sum with wide accumulator	uint8 → NumKong (avoids uint8 overflow); float32 → np.sum(dtype=float64)

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

LUT (lookup tables)

Function	Signature	What it does	How it works
apply_uint8_lut	(img, lut, *, inplace=False)	Apply uint8→uint8 LUT; lut shape (256,) or (C, 256)	Shared (256,): StringZilla or cv2.LUT by size heuristic. Per-channel (C, 256): single cv2.LUT with (256,1,C) table on contiguous HWC; else StringZilla per channel
sz_lut	(img, lut, inplace=True)	Apply shared (256,) uint8 LUT via StringZilla translate	Raw byte translation — channel-unaware, fastest for small images and single-channel

Geometric / spatial

Function	Signature	What it does	How it works
hflip	(img)	Mirror left-right	cv2.flip(img, 1); chunked for >512 channels
vflip	(img)	Mirror top-bottom	cv2.flip(img, 0) for ≤4 channels; NumPy slice for >4 channels
median_blur	(img, ksize)	Median filter (odd ksize ≥ 3)	cv2.medianBlur; ksize ≥ 7 falls back to chunk processing for >4 channels. Accepts float32 via @uint8_io
matmul	(a, b)	Matrix multiply (a @ b)	NumPy @ (BLAS-backed); replaces cv2.gemm which lacks uint8 support
pairwise_distances_squared	(points1, points2)	Squared Euclidean distance matrix (N, M)	Small (N*M < 1000) → NumKong cdist; large → NumPy vectorized ‖a‖²+‖b‖²−2(a·b)

Type conversion

Function	Signature	What it does	How it works
to_float	(img, max_value=None)	Convert to float32 in [0, 1]	float32 → no-op; uint8 → cv2.LUT (256-entry float32 table); others → NumPy divide
from_float	(img, target_dtype, max_value=None)	Scale float32 → integer dtype (round + clip)	float32 → NumPy rint(img * max_value) then clip; non-float32 → generic NumPy path

Decorators (re-exported)

Decorator	What it does
float32_io	Wrap a function: cast input to float32, cast output back to original dtype
uint8_io	Wrap a function: cast input to uint8, cast output back to original dtype

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

Batch Processing

Many functions in Albucore support batch processing out of the box. The library automatically handles different input shapes:

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

Functions will preserve the input shape structure, applying operations efficiently across all images/slices in the batch.

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 different backends (NumPy, OpenCV, custom code).
2. Automatic Selection: The library automatically chooses the fastest implementation based on the input image type, size, and number of channels.
3. Optimized Algorithms: Custom implementations are optimized for specific use cases, often outperforming general-purpose libraries.
4. NumKong: SIMD blend for add_weighted; cdist for small pairwise_distances_squared; wide-accumulator moments for uint8 global mean/std in stats; scale for float32 multiply_by_constant (see docs/numkong-performance.md).

Micro-benchmarks vs NumPy/OpenCV: see benchmarks/README.md; quick run: uv run python benchmarks/benchmark_numkong.py (OpenCV rows are skipped if cv2 is not installed).

See docs/performance-optimization.md for detailed performance guidelines and best practices.

Documentation

• CLAUDE.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/numkong-performance.md - NumKong vs OpenCV/NumPy/LUT baselines (benchmark tables; sum/mean/std)
• docs/public-api.md - Star-exported routers vs albucore.functions shims
• benchmarks/README.md - Python micro-benchmarks (uv run python benchmarks/…)
• docs/research/ - Research notes (extra benchmark writeups; see benchmarks/README.md for scripts)

License

MIT

Acknowledgements

Albucore is part of the AlbumentationsX project. We'd like to thank all contributors to AlbumentationsX and the broader computer vision community for their inspiration and support.