Zignal: A zero-dependency image processing library

3 months ago 1

Zignal is a zero-dependency image processing library heavily inspired by the amazing dlib.

This library is in early stages of development and being used internally. As a result, the API might change often.

zig fetch --save git+https://github.com/bfactory-ai/zignal

Then, in your build.zig

const zignal = b.dependency("zignal", .{ .target = target, .optimize = optimize }); // And assuming that your b.addExecutable `exe`: exe.root_module.addImport("zignal", zignal.module("zignal")); // If you're creating a `module` using b.createModule, then: module.addImport("zignal", zignal.module("zignal"));

pip install zignal-processing

This library is used by Ameli for their makeup virtual try on.

const std = @import("std"); const zignal = @import("zignal"); const Canvas = zignal.Canvas; const Image = zignal.Image; const Point2d = zignal.Point2d; const Rgba = zignal.Rgba; const savePng = zignal.savePng; pub fn main() !void { const allocator = std.heap.page_allocator; // Create a 800x600 RGBA image var image: Image(Rgba) = try .initAlloc(allocator, 600, 800); defer image.deinit(allocator); // Create a drawing canvas const canvas: Canvas(Rgba) = .init(allocator, image); canvas.fill(Rgba.white); const red: Rgba = .{ .r = 255, .g = 0, .b = 0, .a = 255 }; const start: Point2d(f32) = .init2d(50, 50); const end: Point2d(f32) = .init2d(750, 100); const center: Point2d(f32) = .init2d(400, 300); canvas.drawLine(start, end, red, 5, .soft); canvas.drawCircle(center, 100, red, 3, .soft); // Save the result to a PNG file try savePng(Rgba, allocator, image, "drawing.png"); }

Initially, the features in this library are the ones required to get the virtual try on for makeup working. However, we hope that it can be a foundation from which we can build a high quality image processing library, collaboratively.

Current features include:

Principal Component Analysis (PCA) with generic SIMD-accelerated implementation
- works on arbitrary-dimensional data (colors, spatial features, etc.)
- automatic color-to-point conversion using reflection
- image-to-points utilities for computer vision applications
- efficient batch processing and reconstruction
color space conversions (RGB, HSL, HSV, Lab, XYZ, Oklab, Oklch, etc.)
matrix operations with common linear algebra functions
singular value decomposition (SVD) ported from dlib
geometry
- unified Point system with SIMD acceleration supporting arbitrary dimensions
- 2D/3D/4D points with convenient accessors (x(), y(), z(), w())
- vector operations (add, subtract, scale, dot product, norm, distance)
- dimension conversion and projection methods
- rectangles and geometric primitives
- projective, affine and similarity transforms
- convex hull algorithms
simple image struct with common operations
- resize
- rotate
- crop
- blur
- sharpen
- views (called sub_image in dlib or roi in OpenCV.)
Canvas drawing API
- lines with variable width and antialiasing
- circles (filled and outlined) with soft edges
- polygons (filled and outlined)
- rectangles with customizable borders
- Bézier curves (quadratic and cubic) with adaptive subdivision
- spline polygons with tension control for soft curved shapes
- multiple drawing modes: fast (hard edges) and soft (antialiased edges)

One of the greatest things about dlib is the large amount of examples illustrating how to use many of that library features. I plan to showcase most of the features of this library as simple HTML/JS + Wasm examples, which can be accessed from here.

Currently, there are examples for:

First of all, this project would not have been possible without the existence of dlib. In fact, the first version of the virtual makeup try on was written in C++ with dlib and Emscripten. However, we decided to give Zig a go, even if that meant rewriting everything. As a result, we have no dependencies now.

Finally, B factory, Inc, which is my employer and graciously agreed to release this library to the public.

Read Entire Article

Zignal: A zero-dependency image processing library

Related

Ask HN: Does VC math still work in the AI era?

AI QA?

Virgin to launch a rival train service through the Channel T...