AMAZING AUGUST SALE: 25% Off Certificates and Diplomas! Limited-time Offer - ends Friday, 6th August 2021

Claim My 25% OFF

An Introduction to Generative Models in Computer Vision

Study the value of generative models in computer vision in predicting complex distributions with this free online course

Publisher: NPTEL
Have you ever wondered how some high-quality images look exactly like actual images? This course illustrates the use of generative models to train data to make them resemble a real distribution. Learn how generative networks capture correlations on complicated distributions to create similar images in appearance to the original ones. Study the perceptions behind the functioning of deep generative models in this course.
An Introduction to  Generative Models in Computer Vision
  • Duration

    3-4 Hours
  • Students

  • Accreditation






View course modules


This course illustrates the importance of developing mathematical models for performing various computer vision tasks. It begins by illustrating the notion of different generative models in deep learning. You will study the landscape of machine learning and ascertain the importance of generative models in the scheme of deep learning methods. This section will include the procedure for computing the patterns in a dataset using unsupervised learning methods. You will study how generative adversarial networks (GANs) utilize two different networks to generate new data identical to several data models in the training datasets. In addition to this, the process of training and evaluating various types of generative adversarial networks are explained. You will comprehend the significance of variational autoencoders (VAEs) in deep generative models. The multiple applications and the procedure for training these autoencoders using computer vision techniques are described. 

Variational autoencoders are used in dealing with latent space irregularity. This course will explain this concept and include the process of determining the parameters of the distribution using Gaussian models. You will be taught about the various methods that unify various deep generative models using adversarial networks and variational autoencoders. You will explore the process of mapping the generated images of GAN using VAE to improve the standard. This process comprises the procedure for enhancing the quality of the image by integrating variational autoencoders and generative adversarial networks. The process of normalizing the latent code using an adversarial loss mechanism is disclosed. This includes using classifiers to distinguish between original data and created data anticipated by the generative network. Subsequently, you will study how VAE-GAN outclasses conventional variational autoencoders. Replacing the decoder of VAEs using a GAN discriminator for ascertaining the loss function is revealed. 

Finally, the procedure for enhancing the quality of the image by integrating variational autoencoders and generative adversarial networks is described. Next, the course highlights the various deep generative models that have been successful over the years. You will discover the process of modelling and transforming complex densities using non-linear independent components estimation and real-valued non-volume preserving methods. Following this, the process of converting a simple distribution into a multifaceted one is described. This will include the use of flow based-models to model a probability distribution by influencing the normalising flow. Subsequently, integrating animated alterations with manageable variations in probability distributions using autoregressive flows is described. Lastly, the course explains the various methods of estimating the pixels across its axes and the process of choosing a distribution and its conforming loss function to model the output pixels. Study the fascinating ways of training extremely complicated distributions to predict the forthcoming sequence or patterns using large samples in a dataset.


Start Course Now