“This is the 23rd day of my participation in the First Challenge 2022. For details: First Challenge 2022”

  • πŸŽ‰ statement: as one of the bloggers in AI field, ❀️ lives up to your time and your ❀️

    • 🍊 Note: small white note, thank you for checking
    • 🍊 super sub reconstruction column, a read

    Basic information

    • PULSE: Self-Supervised Photo Upsampling via Latent Space Exploration of Generative Models
    • Github.com/adamian98/p…

    Pulse Environment setup

    • Server: Ubuntu16.04 GTX1060 6G * 8
    • Personal user: CUDA version 9.2; Cudnn 7.6.2;
    • Torch ==1.5.0+ Cu92 TorchVision ==0.6.0+ Cu92

    Pytorch official installation command

    Conda create -n torch15 python=3.8.2sourceActivate Torch15 PIP install Torch ==1.5.0+cu92 TorchVision ==0.6.0+cu92 -f https://download.pytorch.org/whl/torch_stable.html PIP install pandas PIP install requests PIP install scipy = = 1.4.1 PIP Install dlib = = 19.19.0Copy the code
    • Download the pre-training model, I downloaded the good pre-training model to share here:
    Link: https://pan.baidu.com/s/1fJ1qtN2NyeCNr0HnCWriOA extraction code: coolCopy the code

    Using the pre-training model, the test:

    1. Detect the face in the original picture and sample it down to 32×32 size and save it in: Input directory;

    python align_face.py -input_dir dataset/mix
Copy the code

    2, based on 32×32 face micrograph, rebuild to generate 1024×1024 HD face macrograph, save to: runs directory;

    python run.py   
Copy the code

    The effect is as follows:

    Meaning and innovation of pulse articles

    • Refer to the link

    The Duke team developed PULSE, an AI image generation model. PULSE can convert a low-resolution portrait into a clear, lifelike portrait in five seconds.

    It should be noted that what PULSE does is not “restore” low resolution images to high resolution, but output many possible high resolution images. For example, if a user inputs a 16-16 resolution image, PLUSE can output a set of 1024-1024 resolution images.

    The research was presented this month at the Conference on Computer Vision and Pattern Recognition CVPR 2020. The paper is titled PULSE: Sampling from Self-supervised Photographs through Potential Spatial Exploration of Generative Models. Self-supervised Photo Upsampling via Latent Space Exploration of Generative Models).

    Two, method: reduce the size loss method: with the generated image “backward” fuzzy graph, similar to output

    To ensure “correspondence” between the output image and the input image, the researchers applied a “downscaling loss” approach to the PULSE model.

    When the generation network of PULSE model proposed a clear image as the output, the discriminant network would reduce the resolution of the clear image to the same level as the input image. Then the discriminant network compares the similarity between the downscaling loss image and the input image.

    Only when the downscaling loss image has high similarity to the input image, the discriminant network will judge that the clear image of the generated network proposal can be used as the output.

    Three, 40 evaluators participated in scoring, and the MOS score of PULSE model was the highest

    The researchers evaluated PLUSE’s performance with CelebA HQ, a high-resolution face dataset. For comparison, we trained supervised models BICBIC, FSRNET, and FSRGAN using the CelebA HQ dataset.

    All models took 1616 resolution images as input, the BICBIC, FSRNET and FSRGAN models took 128128 resolution images as output, and the PLUSE model took 128128 resolution images and 10241024 resolution images as output.

    In terms of image quality, the PULSE model outperformed other models in generating details such as eyes and lips.

    My own summary of the paper

    • Training data (no need to pair lR-HR image data sets)

    • The training process

    • The reconstruction effect

    • The evaluation index

    • The authors conclude: we have established a new method for image super-resolution and a new problem representation.

    Compared with the traditional CNN monitoring work, this opens up a new approach for super-resolution methods along different orbits. This method is not limited to the specific degradation operator seen in the training process, but always maintains a high perceived quality.

    πŸš€πŸš€ The road to AI is long and arduous

    • 🍊 super sub reconstruction, master and doctor technical exchange

    πŸš€πŸš€ So far the gold digging platform has created the following classic blog post πŸš€πŸš€

    Computer vision field, classic blog post

    • 🍊 Use AI to convert photos of good friends into pencil sketches — 🍊2020 U2Net🍊

    • 🍊 NiceGAN environment to build, style migration (with source) | 【 2020 】 CVPR

    • 🍊 multi-stage progressive image restoration – go rain, denoising, fuzzy – effective tutorial (with source) | 【 2021 】 CVPR

    • 🍊 Image inpainting Based on Deep Learning – A Review

    • 🍊 graduation thesis, academic paper writing basic skills and experience – one read

    • 🍊 LaTeX2021 formula preparation, graphic installation, detailed tutorial, a read

    AI learning, deep learning environment building: an understanding

    • 🍊 # Ubuntu install CUDa11.2 for current users

    • 🍊 # Linux and Windows setup PIP image source – the most practical machine learning library download acceleration setup

    • 🍊 # Specify the GPU to run and train Python programs, deep learning single card, multi-card training GPU Settings

    • 🍊 # Install Pytorch and Torchvision in Cuda10.0 for Linux

    • 🍊 # Read SSH password login, public key authentication login

    • Install JDK 11: configure the JAVA_HOME environment variable

    πŸš€πŸš€ Mexic AI

  • πŸŽ‰ as one of the bloggers with the most dry goods in the field of AI, ❀️ lives up to his time and qing ❀️
  • ❀️ If the article is helpful to you, please like, comment to encourage bloggers every minute of serious creation