There are emblematic film that is filmed in black-and-white (actually, we should say, would be shades of gray), which -at some time – it was thought that we might as well colorizarse and thus have a more real, with images. And although the idea seemed good, the problem is that the images in shades of gray, be it photographs or films, they lost the color information and, therefore, to want to colorizar, to approximate the color. Suppose that we want to colorizar an image in which an actor comes out with a dark suit.
However, we do not know a priori if the suit is dark gray, dark blue or black. We can’t know. Of course, you could go to a museum of Hollywood where you may keep the wardrobe original of a movie and then we could paint the picture so much more correct, more accurate with the color that you actually had. But on other occasions, we cannot know what exact color were the things, the objects, the skin of the protagonists, and so on. The result almost always ended up painting the film with pastel colors.
However, this difficulty seems insurmountable, it has been proposed to be solved using neural networks. This is a way of “removing old images” or “Deoldify”, as you have been given in the call to this technique. The most interesting thing of all this is that this effort in particular not only seems to solve the problem, but that is an effort amateur.
Since then that one need not be an academic or have the support of a company of IA to enter into the theme. What is needed is to know what and how to do it, not to mention work a lot, as did Jason Antic, who describes himself as a type dedicated to the software, currently working on neural networks GAN to make colorization and restoration of photos.
Antic has done the work, in general, what does a team of programmers. Your only help is a GPU card 1080TI and yet, their models take between two and three days to be trained. The description of the design of its neural network for your project Deoldify does not seem to be very accurate, but it is clear that it uses a GAN (Generative Adversarial Network), which considers a number of important criteria in order to work. As it is, the gist of this is that your effort seems to be working very well.
Note that the neural network does not recognize the hand and may think that it is a bag
Before to see some photos restored and colorizadas, it should be understood that there is no human intervention in the process. All that is done by the neural network and that, there is that to consider, at times there is not a correct response to the colorization of old photos where you don’t really know what colors were the things that appeared in the image. It is assumed, however, that this is precisely “the job” of the neural network that learns and one that, without doubt, is a difficult process.
There are many examples on GitHub and you have to admit that these are the best results. The results seem to be in any way “biased” by the author himself but, is not this typical behaviour of the researchers of the IA to submit their articles better than they are?
The object of this work is to create photos that are better, photos that in the past were good, and that it now seeks to renew. Perhaps what is important is that a neural network can make this work, and present interesting results where even the errors are very instructive.
The best thing is that the code can be downloaded, which is under license from MIT. There are those that say that if you use this software, you have to train the network yourself. The coefficients final model will probably be published soon, but for the moment there is more that play, train the model and see which settings are the best.
Coolest-hacks.com and Partners.