https://github.com/vtuanpham/srgan_reimplementation_modified

Re-Implementation of SRGAN with symmetric padding for a better merged final image (Anysized input with 4x upscaled)
https://github.com/vtuanpham/srgan_reimplementation_modified

deep-learning deep-neural-networks fft gan generative-adversarial-network image image-processing kivy kivy-application kivy-framework machine-learning srgan srgan-keras super-resolution upscaling

Last synced: about 1 month ago
JSON representation

Re-Implementation of SRGAN with symmetric padding for a better merged final image (Anysized input with 4x upscaled)

• Host: GitHub
• URL: https://github.com/vtuanpham/srgan_reimplementation_modified
• Owner: vTuanpham
• Archived: true
• Created: 2022-11-07T05:40:44.000Z (almost 2 years ago)
• Default Branch: main
• Last Pushed: 2023-01-11T06:19:13.000Z (almost 2 years ago)
• Last Synced: 2024-09-25T20:20:34.692Z (about 1 month ago)
• Topics: deep-learning, deep-neural-networks, fft, gan, generative-adversarial-network, image, image-processing, kivy, kivy-application, kivy-framework, machine-learning, srgan, srgan-keras, super-resolution, upscaling
• Language: Python
• Homepage:
• Size: 9.45 MB
• Stars: 2
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Keras SRGAN Reimplementation with modified Generator

**Keras Re-Implementation of SRGAN with symmetric padding for a better merged final image (Anysize input with 4x upscaled)**

AI-upscale application with user-interface for anysize input image with 4x upscaled output based on SRGAN with modified-Generator.

Added Symmetric padding for a seamless final merged image, add more conv layers as well as more residual blocks, lay out general workflow for anysize input.

**Modified-Generator**

![image](https://user-images.githubusercontent.com/82665400/206855054-fbce99c8-4ad0-47ee-af15-33ce8958264b.png)

**Original Discriminator**

![image](https://user-images.githubusercontent.com/82665400/206855129-b72f86ca-7e54-4a6e-bec9-442e66f6ca73.png)

**Pipeline**

![image](https://user-images.githubusercontent.com/82665400/206857747-d10e8a11-5b83-4cf0-8f61-74e1910bbd5f.png)

**NOTE**: Only support for size up to max dim of **512** (Since perdiction is using predict_on_batch with max tensor size of (16,64,64,3) (that why we are using CPU for prediction to avoid OORam)) if you are using a more powerful CPU or planning to reimplement the code with (1,16,16,3) size input tensor per splited sub image and vstack them later, you can defined the maximum size input by changing the func computeApporiateSize

![image](https://user-images.githubusercontent.com/82665400/206857420-a514481e-5a0a-42d0-98ea-e8819efbae51.png)