Comparisons of Image Magnification Methods
Last Updated 12/24/19
Comparisons Using Mouseovers
Input Image (lossless 48 bit PNG)
(produced by reducing 0.25X with Box Kernel)
When you place the mouse cursor over one of the numbers around the periphery of the following image,
the image will change to an enlargement made by the method corresponding to the number given in the legend below.
This requires that your web browser support javascript.
Please, allow ample time for all images to load before starting.
Legend
- Original before reduction
- Box
- Bicubic
- SAR's Pseudo LAD (Default Settings)
- SAR's Neural Net (MLP) 4X (Multilayer Perceptron), settings
- SAR's Neural Net (MLP) 2X 2X (Multilayer Perceptron), settings
- Waifu2X2X (Convolutional Neural Network not part of SAR, see waifu2x.booru.pics)
- Img Larger (Not part of SAR, see imglarger.com)
- Big Jpg (Not part of SAR, see bigjpg.com)
- Lets Enhance (Not part of SAR, see letsenhance.io)
- Deep Image (Not part of SAR, see deep-image.ai)
- Image Upscaler (Not part of SAR, see imageupscaler.com)
4X Enlargements
Objective Measurements for Clown Image Enlargements.
Results for Y Channel of YCbCr Color Space
|
RMSE1 |
AAE1 |
PSNR2 |
MSSIM2 |
| 2. Box (Nearest neighbor) |
20.14 |
11.90 |
22.05 |
0.5661 |
|
| 3. Bicubic |
18.46 |
11.34 |
22.80 |
0.5889 |
| 4. Pseudo LAD |
16.89 |
10.14 |
23.58 |
0.6407 |
|
| 5. MLP 4X |
16.35 |
9.812 |
23.86 |
0.6533 |
| 6. MLP 2X 2X |
16.51 |
9.923 |
23.77 |
0.6512 |
| 7. Waifu2X2X |
16.03 |
9.690 |
24.03 |
0.6619 |
| 8. Img Larger |
19.80 |
12.24 |
22.20 |
0.5704 |
| 9. Big Jpg |
16.13 |
9.787 |
23.98 |
0.6583 |
| 10. Lets Enhance |
21.82 |
13.96 |
21.35 |
0.5355 |
| 11. Deep Image |
17.44 |
10.81 |
23.30 |
0.6245 |
| 12. Image Upscaler |
17.24 |
10.19 |
23.40 |
0.6704 |
1 - Smaller is better.
2 - Larger is better.
Image with Prominent Halos Around Edges
The above image was produced by an unknown method, but, I have determined that it is well modelled by a bicubic(-2) reduction.
The SAR's MLP used to make the enlargement below was trained with bicubic(-2) reductions.
Legend
- Box
- Bicubic
- SAR's Pseudo LAD, settings
- SAR's Neural Net (Multilayer Perceptron MLP) 4X , settings
- Waifu2X2X (Convolutional Neural Network not part of SAR, see waifu2X)
- Img Larger (Not part of SAR, see imglarger.com)
- Big Jpg (Not part of SAR, see bigjpg.com)
- Lets Enhance (Not part of SAR, see letsenhance.io)
- Deep Image (Not part of SAR, see deep-image.ai)
- Image Upscaler (Not part of SAR, see imageupscaler.com)
4X Magnifications
Key words:
superresolution, super resolution, super-resolution, image enlargement, image magnification, image resizing, image interpolation,
neural network, multilayer perceptron, MLP, convolutional neural network, CNN, free, freeware, image editor, SAR Image Processor
