独家优惠奖金 100% 高达 1 BTC + 180 免费旋转

SRGAN

Media and Entertainment Industry needs a lot of intelligence to meet consumers’ aspirations. GAN, Generative Adversarial Networks is a special kind of deep learning architecture that solves the most complex problems.

GAN can do the following:-

Although we have achieved decent accuracy in super-resolution of images using various approaches in deep learning, we could not get very fine texture details in super-resolution. SRGAN is very powerful to make the super-resolution with very fine details. It is able to achieve great output even with 4x upscaling factors.

Here is the low-quality image:

Figure 2.37: Low-quality input of SRGAN

If you apply SRGAN, you can get a high-quality image (super-resolution). Here is the output of SRGAN and that is what we need in industry.

Figure 2.38: High-quality output of SRGAN

SRGAN can be used for:

· Making super-resolution of all brand images for media

· Making super-resolution of images taken from a normal camera or phone

· Making high-quality HD movie from a low-quality movie

· Making high-quality footage for news reporting

· Converting old low-resolution images into high-resolution images

· Presenting HD content to customers

· Making a new version/high-resolution version of classical movies

The architecture of SRGAN is presented in the following figure:

Figure 2.39: SRGAN Architecture

The result of SRGAN is shown in the following figure:

It aims at converting Old Black and White Movie into Colored Movie

All of us have seen old movies that are black and white movies. Imagine an angel comes to you and you ask for making your old movies color movies. Suppose you see a documentary on World War II, and you wish to see colored version which is almost impossible.

Here comes DeOldify GAN model which converts grayscale image/video into colored image/video.

Input:

Figure 2.41: Black and white input of DeOldifyGAN

Output:

Figure 2.42: Colorful output of DeOldifyGAN

GAN has two deep learning models:

· Generator

· Discriminator

This GAN can be used for

1. Making colored images of gray scale images for media and movies

2. Making colorful images for old gray scale images

3. Making colorful HD movies from black and white movies

4. Making colorful footage for old news reporting

5. Making a new version/colorful version of historical records

To make it better, we add the new attention layer in both the discriminator and the generator and the spectral normalization. We can use their hinge loss and different learning rates for critic versus generator. But this really made the training a lot more stable. Additionally, the attention layers really made a big difference in terms of consistency of coloration, and general quality.

Self Attention GAN architecture and results are explained in Figure 2.43 and Figure 2.44, respectively:

Figure 2.43: Self-attention GAN

Figure 2.44: Result of self-attention GAN

2.6.2 StackGAN

Can you imagine an angel who can generate an image if you tell something?

Please start imagining it as it is happening now due to StackGAN. Stacked Generative Adversarial Networks (StackGAN) is used to generate 256×256 photo-realistic images if we give a text as an input.

Input: A text

A dog is white with black and is sitting next to a boy

Output: An image

Figure 2.45: Output image of StackGAN

StackGAN can be used for:

1. Converting a comment into an image on social media

2. Converting a fact into a meaningful image

3. Making a document more exciting by adding relevant images

4. Making a great image and video based on the novel

5. Making an email exchange more exciting

6. Making a news report very exciting by adding photos

It breaks the hard problem into subproblems through a sketch-refinement process. It works in two stages:

1. The Stage-I GAN aims at getting the primitive shape and colors of the object based on the given text description, and it outputs the low-resolution image. The Stage-II GAN aims at generating high-resolution images with photo-realistic details by using the low-resolution image and text descriptions as inputs.

2. The Stage II GAN aims at working on defects in Stage-I results and add finer details and hence Stage II can be considered as refinement stage. To improve the quality of outputs, we introduce a novel conditioning augmentation technique that encourages smoothness in the latent conditioning manifold.

The architecture of stackGAN is following:

Figure 2.46: Black and white input of DeOldifyGAN

Outputs of stackGAN are the following:

Figure 2.47: Output of Stack-GAN

2.6.3 Face ageing — Age-cGAN

Many social media companies and entertainment companies would like to see how celebrities or any person would look at a very old age. Social media can offer many apps regarding the age change, and this GAN can be useful for that:

Figure 2.48: Output of Age-cGAN

It can be used for:

· Offering the capability to everyone for getting new looks in various age-group

· Modifying the movie if required

· Making a movie without costume change for actors

GAN architecture can be seen in the following:

Figure 2.49: Architecture Of Age-cGAN

2.6.4 BeautyGAN: Facial Makeup Transfer

People want to make-up like different celebrities. This GAN can be used to transfer make-up from desired make-up style to your face.

Below is the rough idea about how BeautyGAN works It aims to transfer and to translate the makeup style from a given reference makeup face image to another non-makeup one while preserving face identity. Extracting and transferring such local and delicate makeup information is done by these existing style transfer methods.

Please feel free to share your feedback

— Navin Manaswi

Author of AI books and AI Evangelist

Add a comment

What is your view?

Send

Related posts:

Using Optimization to cure Cancer

High doses of radiation can kill cells or prevent them from growing and dividing. This is true for cancer cells and normal cells, but radiation therapy is attractive as a cancer treatment because the…

Cannabis Consumption for Beginners

As a public service for the Greenhorns Of Ganja, we have a few tips for those taking their first steps down what we hope will be a very happy road… If you’re new to cannabis, we typically recommend…

How spirituality and saints shaped South Asian society?

Rumi called love a key to everything in the world. A bridge connecting the whole world. Love and care is the way humanity survived throughout the turbulent time of human history. Spirituality and…