Heyo, I'm learning web optimization. I needed to scale a large image down from over a thousand pixels... ...to 51x72 pixels. I used GIMP, and simply changed the resolution to 72 ppi (naturally, no effect) and the size to height 72 (with the width being constrained to 51 as a result; height was the limiting factor) using the Image>Scale Image Tool: It looks excessively blurry. My friend did it in photoshop, saying he just did "Resize Image" or such: I'm not a graphics designer, and have only worked tangentially with such graphics concepts, however image optimization is important for web development, so it would help a lot if I had insight on why these two images are so radically different. I tried Cubic and Linear interpolation as well, to no effect (I'm guessing those are matrix functions that determine the processing of the image whilst scaling, I have no idea otherwise). Is it likely I'm using the wrong tool? Or is it simply that Photoshop costs $$ thus will produce better results? Thank you in advance for any guesses/answers. I understand if it's too vague to answer.
ppi is more important for printing there are a lot of different scaling algorithms some are better for enlargments other for reduction some are better for working with pixelart other with gradients there are also some new one developed that use artificial inteligence for scaling. In PS you can choose form one of few algorithms not sure about gimp. You can check this for more details– XesenixJun 14, 2019 at 17:22
All image scaling introduces some blur since pixels are interpolated. You can mitigate this by sharpening the image after the scaling. For instance, just using Gimp
Filters>Enhance>Sharpen (unsharp mask) with default values:
What exactly does sharpening do? After it is blurry, you've "lost" the data associated with the higher quality image, right? So how would any algorithm make it look nicer with that data gone? Jun 17, 2019 at 17:49
Ok so scaling is a signal processing thing. Conceptually it works as follows*
Discrete image is converted into a continuous image. This is called rebuilding it is done by convolving the discrete signal by a continuous function.
Here you can see two different examples.
Then the signal is re sampled into a new image, Although the scaling up is shown you can sample any which direction you want.
Now the convolution shape affects how you image looks. And there are many many convolution shapes to choose from. The shape you choose is a trade-off between having a more blurry image and having a sharper image (also called ringing). Sometimes you want one over the other which is why you can choose differently in most editors.
* although there are methods that don't work this way, most processes can be reduced to the basic design of this form. Although steps 1 and two are usually baked into one to save time.
Thank you, I like learning theory behind things. Q: where did you get those pictures from? Did you create them yourself or is there a tutorial on these? Jun 17, 2019 at 17:47
@Seraphendipity I drew them a while back, you can find the same image on 3 different posts. But yeah idea for images of my memory from a document labeled "Lore of a TD" a talk in sigraph ages ago by Tony Apodaca. Offcourse i dont remember what convolutions he used, i just convolved the kernel in mathematica or did it manually dont remember exactly.– joojaaJun 17, 2019 at 19:55