Note
This is one of 199 standalone projects, maintained as part of the @thi.ng/umbrella monorepo and anti-framework.
🚀 Please help me to work full-time on these projects by sponsoring me on GitHub. Thank you! ❤️
Extensible bitmap image convolution, kernel presets, normal map & image pyramid generation. This is a support package for @thi.ng/pixel.
This package contains functionality which was previously part of and has been extracted from the @thi.ng/pixel package.
- Convolution w/ arbitrary shaped/sized kernels, pooling, striding
- Convolution kernel & pooling kernels presets
- Higher order kernel generators (Gaussian, Lanczos)
- Image pooling filters (min/max, mean, adaptive threshold, custom)
- Image pyramid generation (w/ customizable kernels)
- Customizable normal map generation (i.e. X/Y gradients plus static Z component)
Floating point buffers can be processed using arbitrary convolution kernels. The following convolution kernel presets are provided for convenience:
Kernel | Size |
---|---|
BOX_BLUR3 |
3x3 |
BOX_BLUR5 |
5x5 |
GAUSSIAN_BLUR3 |
3x3 |
GAUSSIAN_BLUR5 |
5x5 |
GAUSSIAN(n) |
2n+1 x 2n+1 |
HIGHPASS3 |
3x3 |
LANCZOS(a,s) |
as+1 x as+1 |
SHARPEN3 |
3x3 |
SOBEL_X |
3x3 |
SOBEL_Y |
3x3 |
UNSHARP_MASK5 |
5x5 |
Custom kernels can be defined (and code generated) using an array of
coefficients and a given kernel size. See above presets and
defKernel()
for
reference.
Furthermore, convolution supports striding (i.e. only processing & keeping every nth pixel column/row, aka downscaling) and pixel pooling (e.g. for ML applications). Available pooling kernel presets (kernel sizes must be configured independently):
Kernel | Description |
---|---|
POOL_MEAN |
Moving average |
POOL_MAX |
Local maximum |
POOL_MIN |
Local minimum |
POOL_NEAREST |
Nearest neighbor |
POOL_THRESHOLD(bias) |
Adaptive threshold |
Convolution can be applied to single, multiple or all channels of a
FloatBuffer
. See
convolveChannel()
and
convolveImage()
.
See ConvolveOpts for config options.
import { floatBufferFromImage, FLOAT_RGB, imageFromURL } from "@thi.ng/pixel";
import { convolveImage, SOBEL_X } from "@thi.ng/pixel-convolve";
// convolutions are only available for float buffers (for now)
const src = floatBufferFromImage(await imageFromURL("test.jpg"), FLOAT_RGB);
// apply horizontal Sobel kernel preset to all channels
// downscale image by factor 2 (must be integer)
// scale kernel result values by factor 4
const dest = convolveImage(src, { kernel: SOBEL_X, stride: 2, scale: 4 });
Normal maps can be created via normalMap()
. This function uses an adjustable
convolution kernel size to control gradient smoothness & details. Result X/Y
gradients can also be scaled (uniform or anisotropic) and the Z component can be
customized to (default: 1.0). The resulting image is in FLOAT_NORMAL
format,
using signed channel values. This channel format is auto-translating these into
unsigned values when the image is converted into an integer format.
Step | Scale = 1 | Scale = 2 | Scale = 4 | Scale = 8 |
---|---|---|---|---|
0 | ||||
1 | ||||
2 | ||||
3 |
import { ARGB8888, FLOAT_GRAY, floatBufferFromImage, imageFromURL } from "@thi.ng/pixel";
import { normalMap } from "@thi.ng/pixel-convolve";
// read source image into a single channel floating point buffer
const src = floatBufferFromImage(await imageFromURL("noise.png"), FLOAT_GRAY);
// create normal map (w/ default options)
// this results in a new float pixel buffer with FLOAT_RGB format
const nmap = normalMap(src, { step: 0, scale: 1 });
// pixel lookup (vectors are stored _un_normalized)
nmap.getAt(10, 10);
// Float32Array(3) [ -0.019607841968536377, -0.04313725233078003, 1 ]
// convert to 32bit packed int format
const nmapARGB = nmap.as(ARGB8888);
STABLE - used in production
Search or submit any issues for this package
yarn add @thi.ng/pixel-convolve
ESM import:
import * as pc from "@thi.ng/pixel-convolve";
Browser ESM import:
<script type="module" src="https://esm.run/@thi.ng/pixel-convolve"></script>
For Node.js REPL:
const pc = await import("@thi.ng/pixel-convolve");
Package sizes (brotli'd, pre-treeshake): ESM: 2.27 KB
Note: @thi.ng/api is in most cases a type-only import (not used at runtime)
Three projects in this repo's /examples directory are using this package:
Screenshot | Description | Live demo | Source |
---|---|---|---|
Interactive image processing (adaptive threshold) | Demo | Source | |
2.5D hidden line visualization of digital elevation files (DEM) | Demo | Source | |
Normal map creation/conversion basics | Demo | Source |
If this project contributes to an academic publication, please cite it as:
@misc{thing-pixel-convolve,
title = "@thi.ng/pixel-convolve",
author = "Karsten Schmidt",
note = "https://thi.ng/pixel-convolve",
year = 2021
}
© 2021 - 2024 Karsten Schmidt // Apache License 2.0