Ji Changxin (Redy) Good morning, boss!

sundry

home page  »  file  »  sundry  »  Principle analysis and welfare of "undressing" of Deepnude algorithm

Principle analysis and welfare of "undressing" of Deepnude algorithm

Deepnude algorithm "take off" clothes - welfare bar http://fulibus.net/deepnude.html
First, the document is as shown in the figure
Deep learning computer vision (guess)
 
Image Inpainting
 
You can refer to the NVIDIA paper to use partial convolution and fill based on partial convolution to repair the image of irregular holes.
Partial conversion of paper code. Address: https://arxiv.org/abs/1804.07723 and https://arxiv.org/abs/1811.11718
Test in Image_Inpainting experiment( https://api.isoyu.com/Deepnude/Image_Inpainting (NVIDIA_2018). mp4) In the image interface of the video, you only need to use tools to simply smear unwanted content in the image. Even if the shape is very irregular, NVIDIA's model can "restore" the image very realistically. The picture is filled with smeared blanks. It can be described as a one button P picture, and "no ps trace". This research is based on the team of Nvidia's Guilin Big Brother Liu. Ji Changxin also keeps an eye on them, They released a way to edit images or reconstruct damaged images Depth learning method, even if the image has a hole or lost pixels. This is the most advanced method in 2018.

Pix2Pix (pairing data required)

Paper reference: https://arxiv.org/abs/1611.07004
The following is the output after training 200 epochs of Pix2Pix model.

CycleGAN (no pairing data required)

CycliGAN uses a cyclic consistency loss function to achieve training without pairing data. In other words, it can transform from one domain to another without one-to-one mapping between the source domain and the target domain. This opens the possibility of performing many interesting tasks, such as photo enhancement, image coloring, style transfer, etc. Only source and target data sets are required.
Reference Papers https://arxiv.org/abs/1703.10593
 
 
 
Learn more https://github.com/tensorflow/docs/blob/master/site/en/r2/tutorials/generative/cyclegan.ipynb

Use Procedure of DeepNude for Windows

DeepNude can really achieve the goal of image to image, and the generated image is more realistic.
 
 
Ps: Delete the color.cp36-win_amd64.pyd file in the deepnude root directory, and then add color.py( https://api.isoyu.com/Deepnude/color.py )File to get the advanced version of deepnude.

yes Personal suggestions for deepnude to get off the shelf and get back on the shelf

1. Dimensions. Including 156M DeepNude_Windows_v2.0.0.zip and 1.90G pyqtlib.rar;
2. Speed. It takes 30 seconds to convert the image;
3. Content. The image to image neural network is used to automatically remove clothes from women to reveal their nudity. This application is suitable for error application of deep learning.
*DeepNude can be implemented using Tensorflow and model compression technology.
*DeepNude should change the current practice of not respecting women.
summary
In fact, Image to Image is not required. We can use GAN to generate images directly from random values or from text.
 
Obj-GAN: https://github.com/jamesli1618/Obj-GAN
StoryGAN: https://github.com/yitong91/StoryGAN
 
The new AI technology developed by Microsoft Research AI can understand natural language description, draw sketches, synthesize images, and then refine details according to individual words provided by sketch framework and text. In other words, this network can generate images of the same scene according to the text description describing the daily scene.
The current optimal text to image generation model can generate realistic bird images based on single sentence descriptions. However, the text to image generator is far more than just generating a single image for a sentence. Given a multi sentence paragraph, generate a series of images, each image corresponds to a sentence, and visualize the whole story completely.
more https://github.com/yuanxiaosc Follow is right
fabulous ( thirty-six )

This article is written by Ji Changxin Author, article address: https://blog.isoyu.com/archives/deepnude.html
use Knowledge Sharing Attribution 4.0 International License Agreement. Unless the reprint/source is indicated, they are all original or translated by this website. Please sign your name before reprinting. Last editing time: August 1, 2019 at 02:37 p.m

Popular articles

Comments:

4 comments, 0 visitors, 0 bloggers
  1.  tourist
    tourist Published on:

    It seems that it can't be downloaded now

  2.  Bozhou
    Bozhou Published on:

    Boss, thanks for sharing

  3.  SangSir
    SangSir Published on:

    Boss, after reading your article at Welfare Bar, hhhhh, come back and make a card

    •  Ji Changxin
      Ji Changxin Published on:

      Good morning, boss

Post reply

[Required]

I am a human?

Please wait three seconds after submission to avoid unsubmission and repetition