mirror of https://github.com/vladmandic/human
30 lines
1.2 KiB
TypeScript
30 lines
1.2 KiB
TypeScript
|
import * as tf from '../../dist/tfjs.esm.js';
|
||
|
|
import { Tensor } from '../tfjs/types';
|
||
|
|
|
||
|
|
function get1dGaussianKernel(sigma: number, size: number): Tensor {
|
||
|
|
// Generate a 1d gaussian distribution size numbers long
|
||
|
|
const range = tf.range(Math.floor(-size / 2) + 1, Math.floor(size / 2) + 1);
|
||
|
|
const distribution = tf.pow(tf.exp(range.div(-2.0 * (sigma * sigma))), 2);
|
||
|
|
const normalized = distribution.div(tf.sum(distribution)) as Tensor;
|
||
|
|
return normalized;
|
||
|
|
}
|
||
|
|
|
||
|
|
function get2dGaussianKernel(size: number, sigma?: number): Tensor {
|
||
|
|
// This default is to mimic opencv2.
|
||
|
|
sigma = sigma === undefined ? 0.3 * ((size - 1) * 0.5 - 1) + 0.8 : sigma;
|
||
|
|
const kerne1d = get1dGaussianKernel(sigma, size);
|
||
|
|
return tf.outerProduct(kerne1d, kerne1d);
|
||
|
|
}
|
||
|
|
|
||
|
|
export function getGaussianKernel(size = 5, channels = 1, sigma?: number): Tensor {
|
||
|
|
return tf.tidy(() => {
|
||
|
|
const kerne2d = get2dGaussianKernel(size, sigma);
|
||
|
|
const kerne3d = channels === 3 ? tf.stack([kerne2d, kerne2d, kerne2d]) : kerne2d;
|
||
|
|
return tf.reshape(kerne3d, [size, size, channels, 1]);
|
||
|
|
});
|
||
|
|
}
|
||
|
|
|
||
|
|
export function blur(image: Tensor, kernel: Tensor, pad: number | 'valid' | 'same' = 'same'): Tensor {
|
||
|
|
return tf.tidy(() => tf.depthwiseConv2d(image, kernel, 1, pad));
|
||
|
|
}
|