GANs - Generative Adversarial Network

Due to current technological advancement now a days deep learning models are being developed to automate the detection of diseases using medical image like X-Ray, CT Scan images, MRI images and other type of medical images. As we know that for training any deep learning or machine learning model we require a lot of data and collecting that data manually is a time consuming task specially for the medical images. Here GANs come to rescue us as using GANs we can generate the required images synthetically. Generative Adversarial Networks(GANs) are the current state-of-art technology which enables us to generate images which never exists in real. Now let's discuss how GANs works.

GANs is comprises of two types of network one is known as "Generator" and other one known as "Discriminator". Generator takes the random noise either from normal distribution or uniform distribution to generates the image. Discriminator take two inputs one images from the real dataset and second fake images generated by the Generator. The discriminator tries to distinguish between real images and fake images.

In a way we can say that "Generator" acts as forger which tries to fool the "Discriminator". While on other hand we can think of discriminator as a police man who tries to catch the fake images. Both generator and discriminator are trained on alternate basis till the time discriminator will not be able to distinguish the fake images generated by generator from the real images.

Binary cross entropy (BCE) is used as objective function for training of network. During the training both networks update their model parameters on the basis feedback received as value of objective function.

Using GANs we can solve the problem of insufficient training samples for any kind of image classification or object detection task which usually requires the lot of data for the training of the model. Also GANs can be used to solve the class imbalance problem in image classification task as this can help us to generate the images for the minority classes to achieve the better performance of the model.