You don't know anything about p graphs compared to neural networks.

Chestnut from concave non-temple
Quantum bits produced | Public number Qbitai


"Big God, can help me to take the Sheep Honey Street Pat P into my face." Oh yes, not only this, but also like the sheep honey yo ... "

It seems like a difficult thing to take an element out of a photo and mix it with another picture without knowing it. After all, a magical collage effect can be generated in minutes, whether it's manual or Ai computing.

And if "another picture" refers to the human painting, it is even more difficult.

However, Cornell University and the Great Gods of the Derby have produced an algorithm that blends objects into their paintings without any traces of PS.

The efforts of a large number of artists, even the artist himself, have been killed by its poison:

What bike do you want?

 Mona Lisa and Bean

Bigabin

 bought you a Biao.

Fan · Kevin Tsai · High

What algorithm, so heartless.

Remember this name: Deep painterly harmonization.

The existing global stylized algorithms are too weak in the view of the great Gods of Cornell and the P-tease.

They do take care of the entire picture's style migration. It may not be too much of a problem to put together paintings and pictures that are infected.

This is my new outfit.

But as pictured above, the tiny differences between the American shields and the Italian painter Onofrio Bramante appear to be apparent. The performance of global style migration is also relatively modest (such as the left three).

Whether it's removing boundaries, matching colors or refining textures, it's hard to make the paste part of the original style of the painting. we're not the same .

As a result, the great gods felt that they needed to build a locally styled neural network.

The big strategy is to move the characteristic statistics of the relevant parts of the painting (neuron responses) to the corresponding positions of foreign objects--the key is to choose and what should be migrated.

Gatys has a geometrical cat.

Based on the Style migration techniques published by the Gatys team, the small partners built a two-pass algorithm with Vgg and added some stylistic reconstruction losses (reconstruction losses) to optimize the results.

Come on, the algorithm details take a step-by-step look. Step One (first Pass): Rough image Coordination (single scale)

Roughly adjusts the color and texture of exotic elements, and corresponds to parts of the picture that are semantically similar. In each layer of the neural network, the adjacent nearest nerve patches are found, matching the neuron responses of the pasted parts.

roll Forisa, charming eyes.

A step back, the sea wide by diving. First do not need to be too entangled in the quality of the image, because to a certain extent sacrificing quality, the team can design a powerful algorithm to adapt to a variety of painting wind.

Using the gram matrix to calculate the style reconstruction loss, you can optimize the roughly coordinated version.

the consequences of no loss of style

Here is an intermediate product, but the style is very similar to the original painting. Step Two (Second Pass): High quality refinement (multi-scale)

With the foundation of the first step, it is logical to start demanding the quality of the image at this point.

In this step, the fire is concentrated in a middle layer responsible for capturing texture properties, and a corresponding graph (correspondence map) is generated to reject the spatial outliers (spatial outliers).

you're sinking in my Dim canvas

Then, the corresponding graph with spatial consistency is sampled (upsample), which goes into a finer level of neural network. This ensures that, for each output location, the neuron response on all scales comes from the same position in the picture.

As a result, the picture will have a more consistent texture and look much more natural. post-processing

After the baptism of the Two-pass algorithm, the image quality is almost impeccable on the medium to large scale, but there may be less accurate conditions on the fine scale. In other words, can be far from the view and not __ Yan (fill in the blanks: 10 points).

So, the post-processing also takes two steps: chroma noise Reduction

The phenomenon of high frequency distortion, mainly in the Chroma channel, does not have much effect on brightness.

Discovering this feature, the team transforms the image into a cielab color space and, in the guided filter, filters the AB chroma channel with the Luma channel as a guide.

work or life, this is a problem

This method effectively improves the high-frequency color distortion, but it is possible to poke a bigger loophole. And then... Patch Synthesis

There is a second step to make sure that each patch of the output appears on the screen. Use Patchmatch to find a similar patch for each patch. By averaging all the patches that overlap the wind, you can reconstruct the output to ensure that no new content is produced in the picture.

However, the side effect here is to soften the details, and then ask guided filter to divide the image into the underlying (Base layer) and detail layer (Detail layer) to weaken the softening effect.

draw the wind to abuse me times, I, choose dog with local migration curative effect is significant

Experimental results show that it is much better to transfer the collection of some regional feature statistics than to transfer the characteristic statistics of many independent locations. The corresponding diagram of the neural network response is a big favor.

The local style migration algorithm guarantees the spatial consistency and cross-scale consistency in the statistical sense. The quality of the puzzle is flawless, and the two "consistency" exploits are remarkable.

eight degrees of life in golden Arches

After that, the jigsaw can probably abandon the global style migration algorithm. Look at the local, your brain hole can bloom inhuman.

However, I am still fascinated by this 360-degree full-dead-angle Magic P-chart than the imperceptible intrusion.

Universal Copy, universal paste

More acid-cool paper Portal:

Http://www.cs.cornell.edu/~fujun/files/painting-arxiv18/deeppainting.pdf

The authors of this paper include: Fujun Luan (Fo Jun), Sylvain Paris, Eli Shechtman, Kavita Bala and others. They also open up the code and see more shocking examples:

Https://github.com/luanfujun/deep-painterly-harmonization

By the way, Fo Jun, a 2015-year undergraduate degree from Tsinghua University, is currently studying for a PhD in Cornell. He has interned at Facebook, face++, Adobe and other companies.

Art Photo Little brother has been studying the occult of vision.

For example, once known as the "next-generation PS" Deep Photo Style Transfer, is also Fo Jun research results. Deep photo enables pixel-level style migrations.

is also very cool.

One more Thing

If you are temporarily inconvenient to the onlookers on GitHub, we have also engaged in a small AI Master P-graph show. In the quantum bit public number (Id:qbitai) dialogue interface, reply " master " Two words, you can begin to visit.

- End -

Sincere recruitment

The qubits are recruiting editors/reporters to work in the Beijing-Zhongguancun village. Expect talented, enthusiastic classmates to join us. For details, please reply to the "recruit" two words in the quantum bit public number (Qbitai) dialogue interface.

Quantum bit Qbitai Headline number signing author

Վ ' ᴗ ' Ի tracks new trends in AI technology and products