reference:
https://www.jianshu.com/p/8c7a7cb7198c
https://blog.csdn.net/gdymind/article/details/82696481
Zero, Preface
Recently a software called deepnude fire, and the development of its programmer claims that this software can be a key to "off" photos of women's clothes (because of AI training material can only be female), deepnude launched a free version and paid intermediate version, free version of a big watermark consequently covered the basic, paid version just in the upper left corner of the photograph marked "FAKE". The software is a launch set off a great deal of public controversy, most recently the software developers under pressure to turn off the Web site and download the software (an attempt to improve the lives of END) and said the software does not use very deep technology, but pix2pix One application algorithm (based on a large study). Here we will simply introduce pix2pix the AI framework, if you are interested in deepnude itself, then you can focus on the public number "programmer hair loss prevention center" Reply "deepnude" for more information.
First, the inverse picture identification
The effect of deep learning after years of development has been very easy to achieve all recognize AI, previously ridicule had AI workers (data labeling members) main job is to a large number of pictures or data tagging, continuous training AI, reach is "to see a cat is a cat." "look dog is a dog." What do you mean it, is simply to chart a cat can recognize this is a cat, a dog is a diagram to know the dog, by extension is face recognition, license plate recognition, scene recognition and so on, and so on.
Then we will naturally think that it can not put this process in turn, give AI a "cat" to return it to a cat drawing it?
Unfortunately, no! For AI, the cat and know its difficult to draw a cat and Anthropology painting is the same as the same.
Second, the network generates a confrontation GAN
To learn drawing cats teach AI scientists out of the whole GAN, GAN consists of two main elements: G and D, G used to generate random images, D used to determine whether the requirements of this image. Such as the use to continuously generate a random image G, D determines whether the image is a cat FIG. GD interact with each other exercise, if D smart enough, after a period of study G generated images more and more like a cat, you can continually produce the final map cat friends.
Some time ago the fire cat map generator used is such a principle (as well as cartoon picture, beautiful pictures, etc.), but not good enough light to generate a random picture ah, can generate images you need it? For example, standing cat, lying cat, sleeping cat?
Third, paired training pix2pix
GAN is a method used to generate a random picture, and then determine whether it is a cat, the consequences of this model is always random drawing cat you get. Scientists have thought for a an idea, can not give the correct answer first before you judge, let the cat graph G generated closer and closer to the way we need? So there will be paired training pix2pix.
The core algorithm is given two AB pairs of pictures to be trained, as shown below. After extensive training of AI, when we give a X, X AI will automatically be converted to Y.
Theoretically deepnude on a large number of women Pictures (X) and a large number of fruit body picture (Y) conducted on the basis of pix2pix training, and finally to the effect that a given random X, AI automatically generates the results of body Y, of course, the specific operation quite complicated.
IV Summary
Here just pix2pix made a brief introduction, follow-up will continue to add more learning materials, so stay tuned " programmer hair loss prevention center "! If you are interested deepnude itself, then we can focus on the public number "programmer hair loss prevention center" reply "deepnude" for more information.
