Algorithms for most-accurate upscaling image without AI and neural network (gray or color, smooth, rough, furry or contrastBold variants available)
Объяснение принипа работы и больше примеров и сравнений в FullHD: #1, #2
Free and Open Source Image Upscaler
scaleSmooth lets you enlarge and enhance low-resolution images using only math.
Enlarge images and get more quality without losing accuracy (except contrastBoldScale) and hallucinations. It's just math!
scaleSmooth is a windows application on .NET 8.0 on Visual Studio C#, but we'll be glad, if you create your repository of scaleSmooth / scaleRough / scaleFurry / contrastBoldScale / scaleSeparate / scaleBilinearApproximation for other platforms and languages (write me and I'll attach link to your repository) or commit here updates for optimization, batch, interface or more accuracy.
Threshold, autoThreshold, Mean Cuvatute Blur, Median Blur, adjustment by Lanczos, Antiringing and other tools, which can help you get more usefull results for your specific needs are NOT included!
 |
(required Windows 10+ x64, .NET 8.0+)
|
- Just copy necessary function (ScaleSmoothGray / ScaleSmoothColor / ScaleRoughGray and S255 / ScaleRoughColor and S255 / ScaleFurryGray and S255 / ScaleFurryColor and S255 / ContrastBoldScaleGray, S255f and S255 / ContrastBoldScaleColor, S255f and S255 / ScaleSeparateGray, Quadrilateral and Bilinear / ScaleSeparateColor, Quadrilateral and Bilinear / ScaleBilinearApproximationGray, Dist4 and Bilinear / ScaleBilinearApproximationColor, Dist4 and Bilinear) (you can harmless remove ProgressText from code)
- Call it with parameters (image as type Image, int scale, int accuracy - where 0 is fast, 100 is accurate)
- It returns new image as type Image
-
scaleSmooth
Most accurate, but little bit blurred (much less than any interpolation) and mesh structure is still visible
Fast and you can process multiple images at the same time without losing speed -
contrastBoldScale
Perfect result, but too contrasty (for accuracy, subsequent reverse correction is required) and too small details are lost
Fast - Slow, but you can process multiple images at the same time without losing speed -
scaleFurry
Beautiful and detailed result, but only for monochrome images (only pure black and white, or for color - only pure red, black, green, yellow, fuchsia, blue, cyan and white)
Very slow, but you can process multiple images at the same time without losing speed -
scaleRough
Typographic raster stylization, but for monochrome images it gives acceptable result
Slow, but you can process multiple images at the same time without losing speed -
scaleSeparate
Gives monochrome result and there are Gibbs ringing artifacts
Very fast, but you can't process multiple images at the same time without losing speed -
scaleBilinearApproximation
A clearly defined grid structure and Gibbs ringing artifacts are present, but even if these shortcomings are not removed with other tools, it is more accurate than Lanczos and clearer than Lanczos and Bicubic
Fast - Very very slow, and you can't process multiple images at the same time without losing speed
You can increase speed instead of accuracy and vice versa