Introduction
Generative Adversarial Networks (GANs) are the coolest things to have happened to the machine learning industry in recent years. Ever since Ian Goodfellow unveiled GANs in 2014, several research papers and practical applications have come up since and most of them are so mesmerizing that it will leave you in awe for the power of artificial intelligence. In this article, we will go through some really impressive applications of Generative Adversarial Networks (GANs) to blow your mind.
- Also Read – 13 Cool Computer Vision GitHub Projects To Inspire You
- Also Read – 14 Computer Vision Applications Beginners Should Know
- Also Read – 6 Different Types of Object Detection Algorithms in Nutshell
Applications of Generative Adversarial Networks (GANs)
1. Generate Examples for Image Dataset (Data Augmentation)
Sometimes we do not have enough images in the dataset to train the model properly. To diversify the dataset Data Augmentation techniques are used to generate different versions of existing images. Traditionally it used to be done by performing operations like flipping, rotating, etc. But GANs can generate new images from existing in a more advanced manner. The above example shows how GANs can insert sunglasses, smiles, or remove them in the existing images to generate more examples of the image.
GANs majorly work on the principle of generating fake images using the generator network and then passing these images along with real images to the discriminator network for determining how much real like data is generated.
NVIDIA in their recent study was able to depict how this data augmentation technique can be used for augmenting the medical brain CT images comprising different diseases. The researchers used two different methods of data augmentation to evaluate the performance delivered by GANs. In the case of classical data augmentation, the classification performance resulted in 78.6% sensitivity and 88.4% specificity. On the other hand, when GANs were used for synthetic data augmentation, there was an increase in sensitivity and specificity, reaching 85.7% and 92.4% respectively.
2. Generate Photographs of Human Faces
GANs are quite successful in showcasing its ability to produce images of realistic-looking faces of people who do not exist. This is one of the most intriguing applications of Generative Adversarial Networks. All the faces that you see in the example do not exist in reality. This GAN model was trained with the help of some celebrity faces as well, thus some of the faces were quite familiar. These methods were useful in generating objects and scenes as well.
[adrotate banner=”3″]
3. Image-to-Image Translation
Another impressive application of Generative Adversarial Networks is the image-to-image translation which is generally implemented through StyleGAN by using pix2pix approach. An important example of image-to-image translation includes the translation of semantic images to real photographs of buildings and skyscrapers. It can also convert satellite photographs to Google Maps. In addition to all of this, old black and white photographs and even sketches can be converted to colored photographs.
4. Text-to-Image Translation
Another very innovative application of GAN is the generation of photo-realistic images from the text with the help of Stacked GANs. Currently, this GAN can produce photographs of birds and flowers by using text. The future prospect of this application may include describing various types of objects, building/designing a product.
5. Photograph Editing
In this application of GANs, we can edit the photo not just by changing colors or filter but we can change the complete characteristics of the photo. This is achieved with the help of Conditional GAN. In the above example, we can see how GAN has changed hair color, style, facial expression, and also the gender of the person in the image.
6. Photo Blending
Photo Blending means merging features/components of two different pictures for creating a new image. Huikai Wu in 2017 proposed Gaussian-Poisson Generative Adversarial Network (GP-GAN) for blending photos. This practical applications of GANs have been highly impressive to place elements like fields, rivers, large structures/houses are now placed in different images to produce unbelievable results
7. Face Swapping (Deep Fakes)
One of the most talked-about applications of GANs is Face Swapping. In this application, GANs are able to swap the faces of two different people with utter perfection. Face Swapping has been producing some of the most hilarious and controversial results with faces of celebrities swapped on various kinds of inappropriate results.
An enhanced version of face swap is Deepfake, it is a technique where a person’s face can be replaced in a video with another person. Still, images are now able to show emotions and speak about topics that they never said. All of this has been possible because of GAN powered Deepfakes. FaceApp and ZAO are two of the most famous applications that can show how Deepfake and face-swapping actually works.
8. Video Prediction
Predicting what will happen next seems supernatural. But GANs have made this quite natural. With the help of these generative adversarial networks, predicting future frames in videos is now possible. The recent dual motion GAN architecture has learned to enforce the future frame predictions in line with pixel-wise video flow.
9. Privacy-Preserving
Generative Models can be helpful in information sharing. This generally includes crucial pattern-based data, detailed information, and other shape-related data as well. Since GANs use an architecture where they have to learn to differentiate between real and fake samples, they are now trained to identify cyber attacks and malicious viruses. A new technique known as SSGAN (Secure Steganography Based on Generative Adversarial Networks) is used to apply steganalysis over images and perform detection of hazardous encodings that may prove destructive.
10. Domain Adaptation
Usually, we train our model over a specific kind of data (obtained from the real-world). Now consider we are working with computer vision problems where different surrounding lighting, camera angles, or atmospheric conditions can cause the model to falter. These instances show how crucial domain adaptation can be and this problem is addressed with the help of GAN where they can work upon a specific area of an image or help in improving the quality of an image for better results. In the above youtube video, you’ll see how GAN adapts for nighttime vehicle detection.
There are also use cases where images are converted to emojis, translating the colors of a particular component in an image.
11. Drug Discovery
Insilico is one of the leading companies that have leveraged GANs for performing research in finding out the main reason behind a disease by studying the various types of proteins that may act as the driver. The GAN-based models have shown promising performance and have helped in reducing the time consumption and human involvement as well.
- Also Read – 13 Cool Computer Vision GitHub Projects To Inspire You
- Also Read – 14 Computer Vision Applications Beginners Should Know
- Also Read – 6 Different Types of Object Detection Algorithms in Nutshell
Conclusion
In this article, we look at various mindblowing applications of Generative Adversarial Networks (GANs) that would have surely left you impressed. We encourage you to go through the research papers of each of them yourself for in-depth understanding.
