🔥 PNG Upscale - AI Super Resolution 🔥

Small tool using pretrained models to upscale images

Download is available from the Releases Page or Google Drive or MediaFire

• Hosted Folder include full "Models" folder 📁 and executable Files 🖼️ to download

• Windows 64bit [Exe] / [Jar]

• Linux 64bit [Jar]

• macOS 64bit [Jar]

• Download the executable corresponding with your operating system, and the Models folder

• It's possible to download only some of the models if you want (It just wont let you use them inside the program)

• The Models folder needs to be in the same directory as the Jar/Exe to use them

---

• ⚠️ Be careful when trying to resize very large pictures, it can take considerable time and resources ⚠️
• To upscale an image you just need to choose a mode, load a picture and press start
• Save button can be used to choose an output folder and filename before you start the process (either just name or .png)
• You can double click the text box to change [Dark <-> Light] theme (disabled when upScaling)
• Use PNG images for best results
• if faced with a JNI Error see this issue for a possible fix #33

---

Models

all of the model download links below are already included in the MediaFire folder.

There are four trained models integrated into the program :

EDSR

[Best Quality]+[Slowest]

• Size of the model: ~38.5MB x3. This is a quantized version, so that it can be uploaded to GitHub. (Original was 150MB each.)
• This model was trained for 3 days with a batch size of 16
• Link to implementation code: https://github.com/Saafke/EDSR_Tensorflow
• x2, x3, x4 trained models available
• Advantage: Highly accurate
• Disadvantage: Slow and large filesize
• Speed: < 3 sec for every scaling factor on 256x256 images on an Intel i7-9700K CPU.
• Original paper: Enhanced Deep Residual Networks for Single Image Super-Resolution [1]

Trained models can be downloaded from here.

ESPCN

[Fast]

• Size of the model: ~100kb x3
• This model was trained for ~100 iterations with a batch size of 32
• Link to implementation code: https://github.com/fannymonori/TF-ESPCN
• x2, x3, x4 trained models available
• Advantage: It is tiny and fast, and still performs well.
• Disadvantage: Perform worse visually than newer, more robust models.
• Speed: < 0.01 sec for every scaling factor on 256x256 images on an Intel i7-9700K CPU.
• Original paper: Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network [2]

Trained models can be downloaded from here.

FSRCNN

[Fast]

• Size of the model: ~40KB x3
• This model was trained for ~30 iterations with a batch size of 1
• Link to implementation code: https://github.com/Saafke/FSRCNN_Tensorflow
• Advantage: Fast, small and accurate
• Disadvantage: Not state-of-the-art accuracy
• Speed: < 0.01 sec for every scaling factor on 256x256 images on an Intel i7-9700K CPU.
• Notes: FSRCNN-small has fewer parameters, thus less accurate but faster.
• Original paper: Accelerating the Super-Resolution Convolutional Neural Network [3]

Trained models can be downloaded from here.

LapSRN

[Has x8]

• Size of the model: between 1-5Mb x3
• This model was trained for ~50 iterations with a batch size of 32
• Link to implementation code: https://github.com/fannymonori/TF-LAPSRN
• x2, x4, x8 trained models available
• Advantage: The model can do multi-scale super-resolution with one forward pass. It can now support 2x, 4x, 8x, and [2x, 4x] and [2x, 4x, 8x] super-resolution.
• Disadvantage: It is slower than ESPCN and FSRCNN, and the accuracy is worse than EDSR.
• Speed: < 0.1 sec for every scaling factor on 256x256 images on an Intel i7-9700K CPU.
• Original paper: Deep laplacian pyramid networks for fast and accurate super-resolution [4]

Trained models can be downloaded from here.

---

Benchmarks

Comparing different algorithms. Scale x4 on monarch.png

	Inference time in seconds (CPU)	PSNR	SSIM
ESPCN	0.01159	26.5471	0.88116
EDSR	3.26758	29.2404	0.92112
FSRCNN	0.01298	26.5646	0.88064
LapSRN	0.28257	26.7330	0.88622
Bicubic	0.00031	26.0635	0.87537
Nearest neighbor	0.00014	23.5628	0.81741
Lanczos	0.00101	25.9115	0.87057

---

As a Demo this image was resized from 256x256 to 85x85, and then upscaled using this program

x2 Demo (85x85 -> 170x170)

Original	Bicubic Interpolation	EDSR

ESPCN	FSRCNN	LapSRN

Bicubic Interpolation is the standart resizing technique used by most editing tools like photoship etc..

x4 Demo (85x85 -> 340x340)

Original	Bicubic Interpolation	EDSR

ESPCN	FSRCNN	LapSRN

References

[1] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee, "Enhanced Deep Residual Networks for Single Image Super-Resolution", 2nd NTIRE: New Trends in Image Restoration and Enhancement workshop and challenge on image super-resolution in conjunction with CVPR 2017. [PDF] [arXiv] [Slide]

[2] Shi, W., Caballero, J., Huszár, F., Totz, J., Aitken, A., Bishop, R., Rueckert, D. and Wang, Z., "Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network", Proceedings of the IEEE conference on computer vision and pattern recognition CVPR 2016. [PDF] [arXiv]

[3] Chao Dong, Chen Change Loy, Xiaoou Tang. "Accelerating the Super-Resolution Convolutional Neural Network", in Proceedings of European Conference on Computer Vision ECCV 2016. [PDF] [arXiv] [Project Page]

[4] Lai, W. S., Huang, J. B., Ahuja, N., and Yang, M. H., "Deep laplacian pyramid networks for fast and accurate super-resolution", In Proceedings of the IEEE conference on computer vision and pattern recognition CVPR 2017. [PDF] [arXiv] [Project Page]