Image Representation

In Iris, images are represented by the Image<B> struct, where B is a Burn Backend type.

An Image wraps a 3D Burn tensor with the shape format [Channels, Height, Width] containing float values in the range [0.0, 1.0].

Creating Images

You can create an Image from a raw Burn Tensor or by loading an image file.

use burn::backend::wgpu::Wgpu;
use iris::prelude::*;

fn main() -> Result<()> {
    type Backend = Wgpu;
    let device = Default::default();

    // Create a mock image tensor [3 channels, 480 height, 640 width]
    let flat_data = vec![0.5f32; 3 * 480 * 640];
    let tensor = Tensor::<Backend, 3>::from_data(
        TensorData::new(flat_data, [3, 480, 640]),
        &device
    );
    let image = Image::new(tensor);

    println!("Created image with dimensions: {:?}", image.shape());
    Ok(())
}

Loading and Saving

Iris uses the image crate internally to handle file decoders.

type Backend = Wgpu;
let device = Default::default();

// Load an image from file
let img = Image::<Backend>::open("input.png", &device)?;

// Save an image to file
img.save("output.png")?;

Shape Properties

Convenient properties are exposed on the Image struct:

Method	Description
width()	Returns the image width (dimension 2 of the tensor).
height()	Returns the image height (dimension 1 of the tensor).
channels()	Returns the number of channels (dimension 0 of the tensor).
shape()	Returns the [C, H, W] shape array.

Color Conversions

// Convert to grayscale
let gray = image.grayscale()?;

// Convert grayscale back to RGB
let rgb = gray.to_rgb()?;

// Convert between color spaces
let hsv = image.rgb_to_hsv()?;
let rgb_back = hsv.hsv_to_rgb()?;

Geometric Transforms

// Resize to new dimensions
let resized = image.resize(640, 480)?;

// Rotate by angle in degrees
let rotated = image.rotate(90)?;

// Flip horizontally
let flipped = image.flip(true, false)?;

// Crop a region (x, y, width, height)
let cropped = image.crop(10, 10, 200, 200)?;

Gaussian Pyramid

// Build a 3-level Gaussian pyramid
let pyramid = image.gaussian_pyramid(3)?;