This problem has been bugging me since forever, I already have solutions for high quality x2 resizing (fractal interpolation or hq2x), but now that I wanted something x4 as big everything gave poor results.

Here's the source:



At x4 it should be a 1024x768 picture, after a LOT of work (including Reshade, ImAnalyzer, Corel Photo Paint, Irfanview...) I managed to go from this:



(usually what you'd get with Bicubic interpolation/wiener filter, etc.)

To this:



The challenge seems hard since for every pixel of the picture 15 new pixels have to be created.

Still not satisfactory, since details that are clearly there in the original picture are gone, what seems to be happening is that in the original picture the tree looks natural and good, while in the resized image it looks like one painted by some artist.

I refuse to believe that this is state of the art resizing, any help is appreciated.

____________________

Thanks for the tip. The source comes from a 1024x768 wallpaper where the tree is the highlight of it but it's very small, and I wanted just the tree at full size, that's what began it all

I managed to find the original source, that wallpaper at a 1600x1200 resolution (there are 1900xsomething resolutions but are obviously stretched), which allows for a slightly bigger source:



Which surprisingly allows for a much better result after some research as paulguy suggested.

However, I still have problems with color correction (my method requires the image to be split into three images with the YCbCr color space, scaling them with different methods and merging them, which produces an "orangish" result), but once I manage to solve that I'll post the result. Thanks everyone.

____________________

Posted by GreyMario

---

Traffic Light: I can't think of any reason why you would do that in that manner. Either I'm tired or there's some genius reasoning behind it.

---

There are two ways of resizing (besides interpolation that makes the ugly first resize of the first post, great for resizing down, though, and beside giant pixels):

1: Add details: So what you like to see just isn't in the picture, the details aren't there, so what this does is invent them! The image is examined and new details are added, when this process is good it gives the illusion of high quality, problem is that it's too noisy at x4 due to all the fake details.

2: Get rid of details: This is amazingly good for simple pictures, think of the filters used in emulators to soften the picture, getting rid of all the noise gives the illusion that the original size of the picture is what you are seeing. The problem is that all realism is lost, and the picture pretty much looks like an artistic drawing.

3: So what you are looking for is a perfect balance of 2 and 3, with a bit of tweaking this is achievable, the problem is that it's applied to the whole picture, so some detailed parts are too softened and other soft parts still have more fake details than necessary.

The solution? Split the image in channels! RGB didn't do much, but YCbCr worked just nicely, I just applied the first method of enlargement to the black and white image that you get, it retains the added details, while the other two got a tweaked method of 2, because we don't care about the details of the picture colors. I merged the images and the result is an image with strange illumination, so I had to apply color correction to it, which gives this, an acceptable result considering it's x3.4 enlargement, but I still need to work on it so the original colors are matched.

____________________

Yeah, huh, I'm about to give up, I'd just like to see how other people's methods looks like, for comparison.

Here are the methods separately:





So, my idea was to get something in the middle of both results, which kept me thinking inside the box, a fresh view to the problem would be appreciated.

____________________

Here's blackhole's idea:



The best result yet, since the colors aren't messed up.

There's still some bleeding, but I think the source has it too, by looking at its red channel we see:



The bright red parts cause it.

Thanks for the idea blackhole89.

____________________

Posted by Ailure

---

At least you got a good result without using those odd vector-based resizing algorithms.

---

Method 2 "Get rid of details" is vector based, so the best result was a mix of vector based, and fractal interpolation.

Anyway, does someone know a program that emulates what a (crappy?) TV does when enlarging pictures? This tree looks gorgeous on my 27° TV at a much bigger scale, so a program emulating the TV's blurring would be interesting to try.

____________________