Resizing causes artifacts on some images

[JulianaGaibler]

I noticed that the resize algorithms (when used with an image that has slight banding) create very strange artifacts. How noticeable they are varies but they lead to way larger files.

For the tests above, I generated grey gradients in Illustrator and Photoshop, resized them with all available Filter-Types (except Nearest) and once through photoshop. Afterwards I extrapolated the levels so the results are clearer.
I also resized the images with OpenCV in C++ (INTER_AREA interpolation), which yielded similar results to what photoshop produced.

I don't know if this is a bug per se, but still wonder if this behaviour could be fixed.

[fintelia]

I don't think anyone would object to better resize algorithms. If I'm not mistaken, the current set produces good results when increasing image size, but not so great when reducing size. How reasonable do you think it would be to add INTER_AREA interpolation to the existing set?

[197g]

The artifacts still looks somewhat suspicious to me, have a look at the full resolution version. There are almost grid-like darkened pixels across all filter types, that doesn't look correct?

Edit: the banding effect in gaussian is also kind of strange.

[JulianaGaibler]

@fintelia Well, it's hard to work through the code itself. Thankfully someone made the effort to decipher what INTER_AREA is actually doing. And as far as I have been able to understand this, it shouldn't be too hard to implement.

But I agree with @HeroicKatora that the results of the current filters seem incorrect.

[Shnatsel]

Could you provide the exact input images and code you've used? I have a hunch about what's going on but I need specific testcases to confirm it.

[JulianaGaibler]

@Shnatsel Here you go. I tried to reproduce the image from above as close as possible. Since I deleted my reproduction repo a while back :(

As far as I can tell, the number of pixels by which the image is scaled down has a big influence on if and what kind of artifacts are created.

[Shnatsel]

Thank you! Since image crate contains some unsafe code, I feared it could be a memory error. I've checked for memory errors with Address Sanitizer and libdiffuzz. Fortunately, they've turned up nothing.

I'll run it through miri just in case. I'm afraid I can't offer much assistance with fixing the code, though.

[okaneco]

This looks like it was fixed in #1408. I used test-a.png and an updated main.rs from the repo in #862 (comment). The artifacts aren't present in test-b.png resizes either.

Program used

use image::imageops::FilterType;
use image::GenericImageView;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let files = ["test-a", "test-b"];
    let filters = [
        ("Triangle", FilterType::Triangle),
        ("CatmullRom", FilterType::CatmullRom),
        ("Gaussian", FilterType::Gaussian),
        ("Lanczos3", FilterType::Lanczos3),
    ];
    if !std::path::Path::new("./result").exists() {
        std::fs::create_dir("./result")?;
    }
    for file in &files {
        for filter in &filters {
            let img = image::open(format!("{}.png", file))?;
            let (width, height) = img.dimensions();
            let resized = img.resize_exact(width - 25, height - 25, filter.1);

            // Increase contrast and brightness so we can see the generated artifacts
            let resized = if *file == "test-a" {
                resized.brighten(115).adjust_contrast(250.0)
            } else {
                resized.brighten(105).adjust_contrast(250.0)
            };

            resized.save(format!("./result/{}-{}.png", file, filter.0))?;
        }
    }

    Ok(())
}

From left to right, the filters are Catmull, Gaussian, Lanczos3, and Triangle. The montages were created with imagemagick convert +append.

Commit before - 2721cd9

Commit after - 6d55953

---

As expected, the resulting file sizes are much smaller.

Before	Size	After	Size
pre-test-a-CatmullRom.png	126884	post-test-a-CatmullRom.png	27054
pre-test-a-Gaussian.png	99363	post-test-a-Gaussian.png	27013
pre-test-a-Lanczos3.png	148180	post-test-a-Lanczos3.png	27047
pre-test-a-Triangle.png	49231	post-test-a-Triangle.png	27066

[197g]

Closing this as fixed.