Description
Is your feature request related to a problem? Please describe. When searching the Internet for convolution kernels, one often finds some which normalize at the end of the calculation with division.
A typical example for a Gaussian blur might look like this:
| 1 4 6 4 1 |
| 4 16 24 16 4 |
| 6 24 36 24 6 |
| 4 16 24 16 4 |
| 1 4 6 4 1 | / 256
To find the equivalent version which doesn't include a divisor step, one would need to laboriously divide each of those integers with the divisor, resulting in something which might look like this:
[0.00390625, 0.015625, 0.0234375, 0.015625, 0.00390625],
[0.015625, 0.0625, 0.09375, 0.0625, 0.015625],
[0.0234375, 0.09375, 0.140625, 0.09375, 0.0234375],
[0.015625, 0.0625, 0.09375, 0.0625, 0.015625],
[0.00390625, 0.015625, 0.0234375, 0.015625, 0.00390625]
Describe the solution you'd like
What would be nice would be if the convolute() functions added another (optional) argument called divisor which defaults to 1.
Describe alternatives you've considered
Noting that my app is a Node.js daemon, I was able to monkey-patch node_modules/@jimp/plugin-color/dist/index.js inside the convolute() function definition like this:
var ksize = (kernel.length - 1) / 2;
x = isDef(x) ? x : ksize;
y = isDef(y) ? y : ksize;
w = isDef(w) ? w : this.bitmap.width - x;
h = isDef(h) ? h : this.bitmap.height - y;
var source = this.cloneQuiet();
// Begin edit ------------------
var divisor = 256; // Imagine instead that the function accepted it as an argument
// End edit --------------------
this.scanQuiet(x, y, w, h, function (xx, yx, idx) {
var value = applyKernel(source, kernel, xx, yx);
// Begin edit ------------------
value[0] /= divisor;
value[1] /= divisor;
value[2] /= divisor;
// End edit --------------------
this.bitmap.data[idx] = this.constructor.limit255(value[0]);
this.bitmap.data[idx + 1] = this.constructor.limit255(value[1]);
this.bitmap.data[idx + 2] = this.constructor.limit255(value[2]);
});
It worked perfectly, normalizing the darkness/lightness of the resulting blur. Without this, it's problematic coming up with a good-looking matrix whose values sum to one (unless you have access to Matlab).
Of course, there appear to be several interfaces other than just @jimp/plugin-color/dist but I didn't use those.