❤️ [Introduction to Deep Learning Project] ❤️ [Hyperfractional Reconstruction]

❤️ original: Ink science AI

❤️ disclaimer: this is a [big talk super cent reconstruction] blog post, non-professional technical article, please big guy tread lightly

❤️ [take you to know] ❤️

  • 💙 Small tricks to win a heart === “magnify her beauty
  • ❤️ super cent reconstruction ======== big talk BISR

🚀 school sister can

❤️ super partial reconstruction – – no dead Angle zoom 💙

Superfractional reconstruction is simply such a process

🔔 Basic Information

  • Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (ICCV, 2021)
  • A practical super-resolution degradation model for deep blind images is designed
  • Arxiv.org/pdf/2103.14…
  • github.com/cszn/BSRGAN
  • The author’s own thesis writing research starting point

🎄 Abstract and paper contributions

  • 🚀 performance indicators of the track volume is not moving, then open up a new track typical study [worth learning]

It is well known that the single image Super-resolution (SISR) method will not perform well if the assumed degradation model is different from the degradation model in the real image. Although some degradation models take into account additional factors, such as blurriness, they are still insufficient to cover the various degradation of real images. In order to solve this problem, this paper proposes to design a more complex but practical degradation model, which consists of random mixing blur, downsampling and noise degradation. Specifically, ambiguity is approximated by convolution of two gaussian kernels with isotropy and anisotropy; The downsamples were randomly selected from the nearest, bilinear, and bicubic interpolations; By adding gaussian noise of different noise levels and using JPEG compression of different quality factors, the camera sensor noise after processing is generated by inverse forward camera image signal processing (ISP) pipeline model and RAW image noise model. To verify the validity of the new degradation model, we train a deeply blind ESRGAN super parser, which is then applied to super parsing of composite and real images with different degradation. Experimental results show that the new degradation model can significantly improve the practicality of the depth superresolver, thus providing a powerful alternative solution for true SISR applications.

🎄 Environment Construction

  • Training code warehouse: https://github.com/cszn/KAIR
  • The test code warehouse: https://github.com/cszn/BSRGAN
  • Install Pytorch and Torchvision in Cuda10.0 for Linux — you can install any version
# PyTorch version 1.4 -- up to 1.8PyTorch 1.4.0 -, version 1.8.1# here I directly create and activate a PyTorch1.8.0 conda independent environment to run this codeConda create -n torch18 python=3.7.6 conda activate torch18 gitclone https://github.com/cszn/KAIR

cd KAIR/
pip install -r requirement.txt

Copy the code

🎄 code testing

❤️ code download

git clone https://github.com/cszn/BSRGAN.git
cd BSRGAN/

Copy the code

❤️ pre-training model placement

❤️ 4x reconstruction test

python main_test_bsrgan.py 

python main_test_bsrgan.py 

7641MiB occupied by GPU

# output:LogHandlers setup! 21-09-06 07:44:18.248: Model Name: BSRGAN 21-09-06 07:44:18.251: GPU ID: 0 [3, 3, 64, 23, 32, 4] 21-09-06 07:44:21.401: Input Path: Testsets /RealSRSet Results_x4 21-09-06 07:44:21.402:1 --> BSRGAN --> X4 --> Lincoln. PNG 21-09-06 07:44:21.775: ---2 --> BSRGAN --> x4--> building.png ...Copy the code

Effect of the sample

❤️2 times reconstruction test

vim main_test_bsrgan.py

python main_test_bsrgan.py 

python main_test_bsrgan.py 

The GPU occupies 4469MiB

# output:LogHandlers setup! 21-09-06 07:46:19.338: Model Name: BSRGANx2 21-09-06 07:46:19.342: GPU ID: 0 [3, 3, 64, 23, 32, 2] 21-09-06 07:46:22.452: Input Path: testsets/RealSRSet PNG 21 /RealSRSet_results_x2 21 /RealSRSet_results_x2 21 /RealSRSet_results_x2 21 /RealSRSet_results_x2 21 /RealSRSet_results_x2 21 /RealSRSet_results_x2 ---2 --> BSRGANx2 --> x2--> building.pngCopy the code

💙 More pictures of benefits

New head. – You’re welcome

This blog post is dedicated to ❤️ ❤️, about BSRGAN code training – one button three links, more on the following…

[Superfractional reconstruction] BSRGAN [ICCV, 2021] Detailed record of training steps

🚀🚀

  • 🎄 If you feel the article is not satisfied, but also want to go further, then you can come to my other columns to see oh ~
    • ❤️ image style conversion – code environment to build combat tutorial [attention can read]!
    • 💜 image repair – code environment construction – knowledge summary combat tutorial [is said to be ok]
    • 💙 super partition reconstruction – code environment construction – knowledge summary solution to how to make white Moonlight clearer
    • 💛 YOLO column, only actual combat, unreasonable image classification [suggested collection]!
    • 🎄 Cuda series Linux installation tutorial
    • 💜 ubuntu18 cuda11.2 graphic tutorials to the current user install | configuration cuDNN8.1 |
    • 💜 Install CUDA10.0 for the current user on the Linux server
    • 💜 Pytorch and Torchvision are installed in Cuda10.0
    • 💜 Linux can install multiple versions of the Cuda? | give me a new server, what I would arrange Cuda
    • 💜 View the versions of CUDA and cuDNN

    💛 may I be like the star jun such as the moon, every night streamer bright

    💜 the so-called beauty: go looking for your white moonlight, I want to play 💜

    ❤️You are my white moonlight, shining and warming me❤️