Add fast blur algorithm #683

torfmaster · 2024-08-13T17:27:44Z

This PR is an alternative to image-rs/image#2302 which adds the corresponding functionality in the image crate. It adds a fast approximation of Gaussian blur to this library.

It is an attempt to solve image-rs/image#986. For the discussion discussion please refer to the PR in image until it is decided where the functionality will go.

ripytide · 2024-08-14T21:38:41Z

Thanks for opening another PR in this repo, I feel it is much better placed here. I don't have much knowledge about the specific blurring algorithm but I am happy to help with the code quality and maybe adding some benchmarks.

cospectrum · 2024-08-19T09:00:30Z

Some beautification is needed

cospectrum · 2024-08-19T08:56:56Z

src/filter/fast_blur.rs

+    channel_num * idx + channel
+}
+
+fn my_fast_horizontal_blur<P: Primitive>(


I renamed this to horizontal_fast_blur_half in the original PR.

cospectrum · 2024-08-19T09:07:07Z

src/filter/fast_blur.rs

+        / (-4.0 * (w_l) - 4.0);
+    let m = f32::round(m_ideal) as usize;
+
+    let mut box_sizes: Vec<usize> = vec![];


Vec::with_capcity(n) or iter + collect

I chose the latter. As n is usually between 3 and 6 (and in our case it is always 3) the effect is little, that's why I optimized this code for readability.

Remove the explicit Vec<usize> type for readability

It's already gone (in the PR to the image crate).

cospectrum · 2024-08-19T09:08:31Z

src/filter/fast_blur.rs

+    channel: usize,
+    channel_num: usize,
+) -> P {
+    let x = min(width as isize - 1, max(0, x)) as usize;


replace with clamp

fixed (again in original PR)

cospectrum · 2024-08-19T09:10:48Z

src/filter/fast_blur.rs

+/// Kovesi, P.:  Fast Almost-Gaussian Filtering The Australian Pattern
+/// Recognition Society Conference: DICTA 2010. December 2010. Sydney.
+pub fn fast_blur<P: Pixel>(
+    image_buffer: &ImageBuffer<P, Vec<P::Subpixel>>,


Replace with Image type alias from crate::definitions and rename arg to image or img

fix this once it is decided that 'fast_blurwill land inimageproc`

cospectrum · 2024-08-19T09:11:10Z

src/filter/fast_blur.rs

+pub fn fast_blur<P: Pixel>(
+    image_buffer: &ImageBuffer<P, Vec<P::Subpixel>>,
+    sigma: f32,
+) -> ImageBuffer<P, Vec<P::Subpixel>> {


Replace with Image type alias

fix this once it is decided that fast_blur will land in imageproc

cospectrum · 2024-08-19T09:33:02Z

src/filter/fast_blur.rs

+                let val = P::from(val).unwrap();
+
+                out_samples[channel_idx(channel, i + j * height, channel_num)] = val;
+                vals[channel] = vals[channel]


The expression is soo long!

I extracted the index into a speaking variable. I hope this makes the code more readable.

cospectrum · 2024-08-19T16:04:24Z

src/filter/fast_blur.rs

+        for j in 0..width {
+            for channel in 0..channel_num {
+                let val = vals[channel] / (2.0 * r as f32 + 1.0);
+                let val = if val < min_value {


Replace entire if {} else if {} else {} with val.clamp(min_value, max_value)

cospectrum · 2024-08-19T16:05:13Z

src/filter/fast_blur.rs

+) -> Vec<P> {
+    let channel_size = width * height;
+
+    let mut out_samples: Vec<P> = vec![P::from(0).unwrap(); channel_size * channel_num];


No need for explicit Vec type

cospectrum · 2024-08-19T16:07:32Z

src/filter/fast_blur.rs

+    /// Test blur doesn't panick when passed 0.0
+    fn test_fast_blur_zero() {
+        let image = RgbaImage::new(50, 50);
+        let _ = super::fast_blur(&image, -1.0);


Add use super::* above

cospectrum · 2024-08-19T16:09:52Z

src/filter/fast_blur.rs

@@ -0,0 +1,163 @@
+use std::cmp::{max, min};


All min/max function calls can be avoided

cospectrum · 2024-08-19T16:22:42Z

src/filter/fast_blur.rs

+    };
+    let w_u = w_l + 2.0;
+
+    let m_ideal = (12.0 * sigma * sigma


Expression is too long, I'm sure there is a way to simplify it

It`s II (5) in https://peterkovesi.com/papers/FastGaussianSmoothing.pdf. It results from inverting the equation of ideal sigma and the one coming from repeated box blurs. I don't know how to further simplify this polynomial as all of the terms are linearly independent (as polynomials).

let m_ideal = 0.25 * (n as f32) * (w_l + 3.0) - 3.0 * sigma.powi(2) * (w_l + 1.0).recip();

Thank you, now the formula fits into one line.

cospectrum · 2024-08-19T16:24:04Z

src/filter/fast_blur.rs

+    use image::RgbaImage;
+
+    #[test]
+    /// Test blur doesn't panick when passed 0.0


fixed already in original PR

torfmaster · 2024-08-19T19:09:23Z

@cospectrum thank you for your review comments. I am going to address them here and in my PR on the image crate. However, as long there is no decision whether the code will land in image or in imageproc I think it makes sense to continue the review in the PR for the image crate.

torfmaster · 2024-09-04T12:39:21Z

I just wanted to give a short status report (I also cross post this comment to #683)

I added a functional test showing (for an example) that the effect of fast_blur is close to the Gaussian blur implementation in this library
I added a benchmark that shows that fast blur is much faster than the current blur implementation
I think I addressed all code review comments

What would be the next steps?

cospectrum · 2024-10-15T14:53:37Z

merged into image

torfmaster · 2024-10-16T06:29:29Z

As the code is now in the image crate I will close this PR.

torfmaster mentioned this pull request Aug 13, 2024

Fast Blur image-rs/image#2302

Merged

7 tasks

add fast blur algorithm

d8771e1

torfmaster force-pushed the feature/fast-blur branch from 607707b to d8771e1 Compare August 13, 2024 17:45

cospectrum suggested changes Aug 19, 2024

View reviewed changes

cospectrum reviewed Aug 19, 2024

View reviewed changes

torfmaster closed this Oct 16, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add fast blur algorithm #683

Add fast blur algorithm #683

torfmaster commented Aug 13, 2024

ripytide commented Aug 14, 2024 •

edited

Loading

cospectrum commented Aug 19, 2024 •

edited

Loading

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 31, 2024

torfmaster Aug 31, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024 •

edited

Loading

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

cospectrum Aug 25, 2024

torfmaster Sep 1, 2024

cospectrum Aug 19, 2024

torfmaster Aug 20, 2024

torfmaster commented Aug 19, 2024

torfmaster commented Sep 4, 2024

cospectrum commented Oct 15, 2024

torfmaster commented Oct 16, 2024

Add fast blur algorithm #683

Add fast blur algorithm #683

Conversation

torfmaster commented Aug 13, 2024

ripytide commented Aug 14, 2024 • edited Loading

cospectrum commented Aug 19, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

torfmaster Aug 20, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

torfmaster commented Aug 19, 2024

torfmaster commented Sep 4, 2024

cospectrum commented Oct 15, 2024

torfmaster commented Oct 16, 2024

ripytide commented Aug 14, 2024 •

edited

Loading

cospectrum commented Aug 19, 2024 •

edited

Loading

torfmaster Aug 20, 2024 •

edited

Loading