Image-augment NPM

image-augment

Augment images (geometric, noise, ...) for visual machine learning data augmentation.

This library has been freely inspired from imgaug

This library is intend to work

with hasard for randomness
with 2 different backends :
- opencv4nodejs
- tensorflowjs

Installation

npm install image-augment

Simple example

// First you need a backend for image processing
// this can be one of the following :
// * @tensorflow/tfjs
// * @tensorflow/tfjs-node
// * @tensorflow/tfjs-node-gpu
// * opencv4nodejs

const tf = require('@tensorflow/tfjs-node');

// Then initialize with the backend

const ia = require('image-augment')(tf);

// Create an augmentation pipeline
const basicAugmentation = ia.sequential([
	// Add a noise with a standard deviation of 15
	ia.additiveNoise(15),
	// Rotate 30°
	ia.affine({rotate: 30}),
	// Add a blur kernel of 3 pixel
	ia.blur(3)
]);

// tensorflow backend needs Tensor4d <-> filename function
// see test/examples/simple-example.js for full implementation of those helpers (fileToTensor and tensorToFile)

fileToTensor('test/data/tfjs/lenna.png')
	.then(({images}) => {
		return basicAugmentation.read({images});
	})
	.then(({images}) => {
		return tensorToFile('test/data/tfjs/lenna-example.png', {images});
	})
	.then(() => {
		console.log('done');
	});

Output is :

Grid example with opencv4nodejs

const h = require('hasard');
const cv = require('opencv4nodejs');
const ia = require('image-augment')(cv);

// Random example images
const sometimes = (aug => h.value([aug, ia.identity()]));

const seq = ia.sequential({
	steps: [
		ia.fliplr(0.5),
		ia.flipud(0.5),
		ia.pad({
			percent: h.array({size: 2, value: h.number(0, 0.1)}),
			borderType: ia.RD_BORDER_TYPE,
			borderValue: h.integer(0, 255)
		}),
		sometimes(ia.crop({
			percent: h.array({size: 2, value: h.number(0, 0.1)})
		})),
		sometimes(ia.affine({
			// Scale images to 80-120% of their size, individually per axis
			scale: h.array([h.number(0.6, 1.2), h.number(0.6, 1.2)]),
			// Translate by -20 to +20 percent (per axis)
			translatePercent: h.array([h.number(-0.2, 0.2), h.number(-0.2, 0.2)]),
			// Rotate by -45 to +45 degrees
			rotate: h.number(-45, 45),
			// Shear by -16 to +16 degrees
			shear: h.number(-16, 16),
			// If borderType is constant, use a random rgba value between 0 and 255
			borderValue: h.array({value: h.integer(0, 255), size: 4}),
			borderType: ia.RD_BORDER_TYPE
		}))
	],
	randomOrder: true
});
const image = cv.imread('test/data/opencv4nodejs/lenna.png');

seq.toGrid({images: [image, image, image, image, image, image, image, image]}, {
	filename: 'test/data/opencv4nodejs/lenna-grid.png',
	imageShape: [300, 300],
	gridShape: [4, 2]
});

Output :

Grid Example with tensorflowjs

const h = require('hasard');
const tf = require('@tensorflow/tfjs-node');
const ia = require('image-augment')(tf);


// Random example images
const sometimes = (aug => h.value([aug, ia.identity()]));

const seq = ia.sequential({
	steps: [
		ia.fliplr(0.5),
		ia.flipud(0.5),
		ia.pad({
			percent: h.array({size: 2, value: h.number(0, 0.1)}),
			borderType: ia.RD_BORDER_TYPE,
			borderValue: h.integer(0, 255)
		}),
		sometimes(ia.crop({
			percent: h.array({size: 2, value: h.number(0, 0.1)})
		})),
		sometimes(ia.affine({
			// Scale images to 80-120% of their size, individually per axis
			scale: h.array([h.number(0.6, 1.2), h.number(0.6, 1.2)]),
			// Translate by -20 to +20 percent (per axis)
			translatePercent: h.array([h.number(-0.2, 0.2), h.number(-0.2, 0.2)]),
			// Rotate by -45 to +45 degrees
			rotate: h.number(-45, 45),
			// Shear by -16 to +16 degrees
			shear: h.number(-16, 16),
			// If borderType is constant, use a random rgba value between 0 and 255
			borderValue: h.array({value: h.integer(0, 255), size: 4}),
			borderType: ia.RD_BORDER_TYPE
		}))
	],
	randomOrder: true
});

// tensorflow backend needs Tensor4d <-> filename function
// see test/helpers/files-to-images for full implementation of those helpers (fileToTensor and tensorToFile)
const filenames = new Array(8).fill('test/data/opencv4nodejs/lenna.png');

filesToImages(filenames, seq.backend).then(images => {
	seq.toGrid({images}, {
		filename: 'test/data/tfjs/lenna-grid.png',
		imageShape: [300, 300],
		gridShape: [4, 2]
	});
})

Output :