What is Iris?

Iris is a fast computer vision library written entirely in pure Rust. It leverages the Burn framework for GPU/CPU tensor computations — memory safe, type safe, with no external C/C++ dependencies.

Key Goals

1. Pure Rust Ecosystem: Keep dependencies clean, safe, and fully compiled under the Cargo toolchain.
2. Familiar CV API: Provide developers with a clean, modern API covering Image I/O, Filters, Gradients, Contour tracking, Cameras, Drawing, and Detection.
3. Deep Learning First: Out-of-the-box loading of ONNX and Safetensors models, powered by Burn backends (WGPU, CUDA, Metal, Ndarray).
4. Multi-Threaded CPU Acceleration: All computationally intensive CPU operations are parallelized using Rayon, allowing high throughput when not running on a GPU.

Project Structure

Iris is designed with modularity in mind:

Module	Description
core	Basic matrix representations (Mat), geometric types (Point, Rect, Size, Scalar), and RNG utilities.
image	Image struct, file I/O, resizing, cropping, flipping, rotation, warping, and remapping.
filters	Box blur, Gaussian blur, median filter, bilateral filter, separable 2D filtering, distance transform, and more.
edges	Canny edge detection, Sobel, Scharr, Laplacian, Hough lines, and Hough circles.
morphology	Dilation, erosion, opening, closing, and custom structuring elements.
threshold	Binary, Otsu, triangle, and adaptive thresholding.
color	RGB/HSV/HLS/XYZ/Lab/YUV/YCrCb conversions, channel splitting and merging.
histogram	Histogram computation, equalization, CLAHE, LUT, and comparison.
drawing	Lines, rectangles, circles, ellipses, text, polylines, arrows, and markers.
contours	Contour detection, convex hull, moments, shape analysis, and matching.
features	ORB feature detection (FAST + BRIEF), keypoints, and BFMatcher.
optical_flow	Lucas-Kanade sparse flow and Farneback dense flow.
tracking	MOSSE tracker, background subtraction, and mean-shift tracking.
photo	Non-Local Means denoising and HDR merge (Mertens).
dnn	ONNX model loading, weight loaders, blob preprocessing, and NMS.
segmentation	Semantic segmentation, connected components with stats, and watershed segmentation.
noise	Gaussian, salt-and-pepper, and speckle noise generation.
camera	Camera capture and camera calibration.
face	Face detection and face recognition.
video	Video capture, reading, writing, frame iteration, and metadata.
ml	K-Means clustering.
inpainting	Image inpainting for restoring damaged or missing regions.
stereo	Stereo vision and depth estimation.
kalman	Kalman filter for state estimation and tracking.
hog	Histogram of Oriented Gradients (HOG) descriptor for object detection.
feature_matching	FLANN-based approximate nearest neighbor matching for feature descriptors.