Generative Adversarial Networks

Part
01
of three
Part
01

Generative Adversarial Networks - Part One

Three additional research papers are "Generative Adversarial Network with Attention for Face Attribute Editing", "Eye In-Painting with Exemplar Generative Adversarial Networks", and "Learning Residual Images for Face Attribute Manipulation" by use of the GAN.

Generative Adversarial Network with Attention for Face Attribute Editing

Link
Results
  • The research concludes that a spatial attention mechanism should be introduced into the GAN framework and thus creating a SaGAN method to ensure that face attribute editing is more accurate.
  • The research paper results show that the use of spatial attention within the attribute specific region has been manipulated.
Pros and Cons
  • One of the benefits of this is that it can help face recognition through data augmentation.
  • Another benefit is that the proposed spatial attention to GAN can perform far much better than existing face attribute editing methods.
  • Another benefit is that one is able to manipulate very specific attributes while keeping the rest of the irrelevant regions unchanged within a specific attribute region.
  • This kind of spatial attention mechanism ensures the manipulation of attributes only within the attribute-specific regions while keeping the rest irrelevant regions unchanged.
  • The research report does not give any cons but says there is a need for more research into general image editing tasks.
Timing Estimates
  • The research paper does not give any timing estimates but says that there is a need for more research into general image editing tasks.

Eye In-Painting with Exemplar Generative Adversarial Networks

Link
Results
  • Based on the Facebook research, while an image has some sort of identifying a feature, exemplar GAN provide a useful solution for image generation or in-painting.
  • The research reveals that exemplar GAN can provide superior perceptual results as they can use identifying information stored in reference images or perceptual codes by incorporating them.
  • The research report reveals that it was able to observe that the in-painting quality is sensitive to mask placement and size.
Pros and Cons
  • One benefit of using exemplar GANs is that they are able to be extended to other tasks within computer vision and other domains since they are a general framework.
  • The report reveals that by using exemplar GAN leads to superior results in other in-painting tasks such as filling in missing regions from a natural but uniquely identifiable scene.
  • The report reveals that a generated eye color that more closely matches the reference image is achievable by assigning a higher-weighted loss to the eye color via iris tracking
  • The research report shows that, in the future, there is a need to try more combinations of code-based and reference-based exemplars.
Timing Estimates
  • The research paper does not give any timing estimates but says that there is a need for more research in the general future to try more combinations of code-based and reference-based exemplars.

Learning Residual Images for Face Attribute Manipulation by use of the GAN

Link
Results
  • The research paper tackles the task of face attribute manipulation by using the Generative Adversarial Network.
  • The research adopted the use of both residual image learning and dual learning to successfully manipulate face images while leaving most details in the attribute areas unchanged.
  • The research was able to allow them to focus on attribute specific areas and then learn from each other through the image transformation networks.
Pros and Cons
  • One of the research benefits was that this was able to manipulate the face attribute according to a given attribute value without manipulating the whole image.
  • Another research benefit is that one is able to distinguish the generated images from real images by using the transformation and discriminative networks responsible for attribute manipulation.
  • The research also reveals that residual images can be effectively learned and used for attribute manipulations as another benefit for this.
  • The research reveals the by using a feed-forward Convolutional Neural Network (CNN) and combining it with Generative Adversarial Networks (GAN) one is able to transform the image.
  • The only drawback is that this has to be used in combination with the feed-forward Convolutional Neural Network (CNN).
Timing Estimates

Your Research Team Applied The Following Strategy

Our initial approach was to look for research articles and research papers detailing the use of Generative Adversarial Networks in use during the changing of facial attributes. We were able to find several research papers that talked about GAN. We then focused on those research reports that put into focus on realistically changing facial attributes by using GAN. With this, we were able to locate three research papers and provided them.
Part
02
of three
Part
02

Generative Adversarial Networks - Part Two

Three additional research papers surrounding state of the art for changing facial attributes on demand in a photo-realistic way are "Generate Identity-Preserving Faces by Generative Adversarial Networks", "From attributes to faces: A conditional Generative Adversarial Network for face generation", and "Generative Adversarial Network Fitting for High Fidelity 3D Face Reconstruction."


Generative Adversarial Networks Research Papers

1) Generate Identity-Preserving Faces by Generative Adversarial Networks

Link

  • Here is a link to the research paper.

Information on Results

  • The research reveals that by fusing the generator of Generative Adversarial Networks (GAN) and FaceNetto, one is able to produce images without blurs.
  • It also concludes by claiming it is the very first time GAN has been utilized to produce identity-preserving faces.
  • Additionally, the research shows a universal framework that can be used in selecting distinct generative models as well as discriminators.

Pros and Cons

  • One of the benefits, per the research report, is that by using trained GANs to generate plausible faces together with FaceNet, one is able to generate identity-preserving faces with high quality.
  • Another benefit is that the research presents a universal framework that can be identified in numerous ways.
  • The research also reveals that if one is provided with a target face image, they should be able to produce different "face images of the same identity with different attributes."
  • The research report does not present any cons, but it says that this is the first time GAN has been used to create identity-preserving faces.

Timing Estimates

  • The research paper does not offer any timing estimates, but it claims that its procedure is qualified to spawn high-quality identity-preserving and plausible faces.


2) From attributes to faces: A conditional Generative Adversarial Network for face generation

Link

  • Here is a link to the research paper.

Information on Results

  • The research report affirms that one can produce realistic faces via attribute labels, basing their work on a DC-GAN, or deep conditional convolutional generative adversarial network, approach.
  • It also reveals that attribute-based face generation is possible by using the deep conditional convolutional generative adversarial network approach.
  • The work is essential since as its related application involves entertainment and law enforcement.

Pros and Cons

  • One of the benefits, per the research report, is that it can be used by law enforcement and the entertainment industry for determining certain resemblances of generated images.
  • Another benefit is that the findings signify a method that can be used to create realistic faces using attribute labels.
  • The research reveals that one benefit is that one can generate realistic faces for witness descriptions in which the descriptions serve as the sole evidence available.

Timing Estimates

  • The research paper does not provide any timing estimates, but it states that work in the future will incorporate studies "related to the matching of generated faces with existing faces."


3) Generative Adversarial Network Fitting for High Fidelity 3D Face Reconstruction

Link

  • Here is a link to the research paper.

Information on Results

Pros and Cons

Timing Estimates

  • The research paper does not give any timing estimates, but it states that the evaluation has exhibited exceptional results for texture reconstruction with high quality.


Research Strategy:

Our initial approach was to look for research articles and research papers detailing the use of Generative Adversarial Networks for the changing of facial attributes. We were able to find several research papers discussing GAN. We then focused on those research reports centered on realistically changing facial attributes through GAN. With this, we were able to locate three relevant research papers.
Part
03
of three
Part
03

Generative Adversarial Networks - Part Three

Some additional research papers or data repositories (repos) surrounding state of the art for changing facial attributes on-demand in a photo-realistic way, using Generative Adversarial Networks (GAN) include "StyleGAN: Use machine learning to generate and customize realistic images," "Conditional CycleGAN for Attribute Guided Face Image Generation," and "High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs."

StyleGAN — Using Machine Learning to Generate and Customize Realistic Images

  • The article is available via this link.

Results

  • The article discusses a new architecture that can separate the high-level characteristics/attributes like a person's identity from their low-level attributes (like their hairstyle) within an image.
  • The article analysis the number of images trained within a given time with the "progressive growing" method as well as the traditional GAN.
  • The method can alter fine styles and details related to an image such as "color of the eyes or other microstructures."
  • Data obtained from "Towards Data Science" shows how two images are generated and then combined by collecting low-level features from one and mixing with high-level features from another. A mixing regularization technique utilized by the generator is also explained. The mixing causes some percentage of both images to appear as the output image.

Pros and Cons

  • As a pro, the article reveals that many security systems utilize similar measures such as facial recognition as well as images from the major part of various data on the web to function.
  • The Generative Adversarial Networks (GAN) discussed by the article is StyleGAN. It is the most potent GAN in the world and can generate "synthesized images from scratch in high resolution."
  • As a con, the article states that some people would dub the capabilities of StyleGAN as "scary." Additionally, such advances in machine learning are making specialized skills in the field of image manipulation and engineering to become redundant.

Timing Estimates

  • The article reveals various training/timing estimates related to machine language and images of multiple pixels. The training time is an estimate that varies based on pixels range from a few days to a few weeks.
  • The maximum time estimated for training based on the number of GPUs used is 34 days and 21 hours for a 512 x 512 image resolution.

Conditional CycleGAN for Attribute Guided Face Image Generation

  • The article is available via this link.

Results

  • The study reveals how GANs can "generate highly realistic images," mainly, SISR also known as the SRGAN. The process has produced impressive results with an up-scaling factor up to four.
  • The hallucinated facial features are detailed and look very realistic. They may not tally or conform to any real person's face.
  • Another GAN, the cycleGAN (CycleGAN2017), is known to address the image-to-image translation problem through the use of unpaired image data. It has produced several state-of-the-art results already. The paper capitalizes on the cycleGAN, and proposes the conditional cycleGAN where a face's image result is generated and "subjected to input face attribute condition."
  • The study also illustrates how a lady's face is changed to look like a man's face whose identity "is given by the light-CNN feature."
  • A method known as enforcing forward-backward consistency is in the study. This method has emerged to be one of the most effective ways of training GAN.
  • The methods described can preserve the SR face identity but vary other attributes such as "mouth is closed," or reveal the smiling mouth that is "open with teeth showing."

Pros and Cons

  • One of the processes described by the study has produced impressive results with an up-scaling factor up to 4.
  • One of the methods analyzed has emerged to be one of the most effective ways of training GAN.
  • Although the hallucinated facial features are detailed and look very realistic, they often do not tally or conform to any real person's face.

Timing Estimates

  • The research paper does not give any timing estimates

High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs

  • The article is available via this link.

Results

  • The article provides a quantitative comparison of leading GAN methods. It also reports a subjective human perceptual study and shows a few examples and results obtained from interactive object editing.
  • The research discusses the implementation details and use of LSGANs for stable training.
  • The study also reports on additional qualitative results through two datasets known as the ADE20K dataset and the Helen Face dataset.
  • The results in the paper reveal that conditional GANs can synthesize high-resolution photo-realistic images without the need for "hand-crafted losses or pre-trained networks." The research elaborates on various techniques used to change the skin color or add eyebrows and beards using the researched interface.
  • The method used by the ADE20K dataset generates images that have a "similar level of realism" when compared to original images.

Pros and Cons

  • The study elaborates methods than can generate photo-realistic images without the need for "hand-crafted losses or pre-trained networks."
  • According to figure 6, which states limited time comparison results, the preference or preferred rate of real images is higher than that of images generated through the researched procedure. The system allows the production of more flexible manipulations for high-resolution "results in real-time." However, the building and editing of virtual environments are quite expensive as well as time-consuming.

Timing Estimates

  • According to figure 6, which states limited time comparison results, compared results are plotted on a scale calibrated between 125 to 8000 milliseconds to the x-axis against a preference rate of 50 to 100% on the y-axis.

Face Aging with Contextual Generative Adversarial Nets

  • The article is available via this link.

Results

  • The results of the research demonstrate a proposed framework to produce appealing results through the comparison of the state-of-the-art with ground truth with an observed "performance gain for cross-age face" veri€cations.
  • The study reveals and discusses the architecture of a proposed C-GANs network shown in Figure 2. The input image x is aligned and parsed then paired with an arbitrary age label y.
  • The result gets fed into a conditional transformation network G for a synthesized face. The age discriminative network judges the synthesized face as real or fake.
  • The GAN-based method is also used to generate automatic face aging images and emphasizes the preservation of the original facial looks of a person/their identity by introducing "Identity-Preserving" optimizations.
  • The technique discussed can learn a face manifold and simultaneously traverse specific "smooth age progression" or "regression." This process changes the facial attributes of a person to look older or younger as required.

Pros and Cons

  • The face aging technology discussed in the research, also known as age progression, attracts many research interests because it has several applications. It can be used across various domains such as cross-age face recognition identification of lost children, as well as in entertainment.
  • Generally, the discussed techniques require "sucient age sequences as the training data," and this limits the practicality of the methods described in the study. Although the presented method has "realistic and noticeable transformations," its limited practicability is a con/disadvantage of the techniques discussed in the study.

Timing Estimates

  • The study does not give the timing estimates required for training or implementation. However, it states that "normalizing the input" may result in the training process converging faster (complete in a shorter interval/duration).

Research Strategy

The study investigated research articles and research papers detailing the use of Generative Adversarial Networks to change facial attributes. The study found several research papers that discussed GAN. It focused on research reports that focus on realistically improving/changing facial attributes through GAN techniques. Articles already covered by the previous studies are excluded, and a few research papers are provided.
Sources
Sources