Research Projects
Deep Generative Models for Image Processing (Master's Thesis)
Guide : Prof. Kaushik Mitra
Most of the current deep learning based approaches for image restoration use feed-forward networks to learn the mapping between corrupted image and clean image. But whenever the type or intensity of degradation is changed, the network architecture needs to be modified and the network parameters need to be re-learned. The main idea in this project is to use deep generative models provide a task agnostic and much more versatile solution to image restoration tasks by modeling the image prior distribution.
Compressive Image Recovery using Recurrent Generative Model
In this work, we use auto-regressive generative model, RIDE, for compressive image restoration tasks such as random mask inpainting and single pixel camera recovery. We utilize RIDE’s ability as an image prior to model long term dependencies for reconstructing compressively sensed images. We use backpropagation to inputs while doing gradient ascent for MAP inference.
- Random Mask Inpainting
| Original Image | Masked Image | During Gradient Ascent | Recovered Image |
|---|---|---|---|
 |
 |
 |
 |
- Single Pixel Camera Recovery
| Original Image | Initial Image | During Gradient Ascent | Recovered Image |
|---|---|---|---|
 |
 |
 |
 |
Generative Colorization of Grayscale Images
In this project, we levarage the application of conditional generative adversarial networks in producing different plausible colorizations of the same grayscale image. We are currently exploring different architectures which enforce stochasticity, so that there is significant variation in the produced colorized images.