A person developed "Google's machine learning system" TensorFlow "for" idle face image generator "

Google is a system for deep learning and machine learning "TensorFlowOpen source as open source, applying this to screen cucumbers with deep learning "Cucumber sorting machineThere are farmers who have developed. Attempts have also been made to combine such TensorFlow and image generation neural net "DCGAN" to generate idle face images from scratch.

Idol face image generation with DCGAN by TensorFlow - paste memo
http://memo.sugyan.com/entry/20160516/1463359395

Generating idle face image with DCGAN by TensorFlow Subsequent experiments etc. -
http://memo.sugyan.com/entry/2016/10/12/084751

It is Mr. Kan Jaan who is doing "generation of idle face image with DCGAN by TensorFlow", and the results of trial and error are posted on the blog. First of all, 120 face images were extracted from 90 idles and the total of 10,800 facial images were learned by deep learning as follows. Although it is surely an idol-like face image is generated, most of it is collapsed and it looks like a spirit photograph.


From this point on, collecting more idle images and using a total of 26,000 face images collected from 260 idols, a total of 26,000 subjects are used as learning data, and this is a after-after image that has been changed and experimented repeatedly. In the after-side image on the right side, you can see that idle at the level which can be said that it is comparable to the real human being is not generated by looking.


In addition, you can see in the following movie how the precision of the photograph that is generated while trial and error are repeated by changing and experimenting improving.

DCGAN with feature matching - YouTube


The blog says, "Even if it seems to be somewhat beautiful to make something that looks like a face, it is still a lot of things that have collapsed.When you observe the change, it will be of some quality around 10,000step It was a feeling that there was not much change from there even if it continued until 30 thousand step. "It seems that there will be less change from a certain line. Therefore, a Web UI which can change various input numerical values ​​was created, and the result after passing further trial and error is as follows. There are many similar hairstyles, but we have succeeded in generating fairly human-sized faces.


In the future, designating elements such as blond hair, short, funny eyes, smile and so on "the goal is to be able to freely create your favorite face" is the goal.