TUTORIAL – HALF DAY SESSION
Image Restoration, Decomposition, Synthesis, and Style Transfer via Unsupervised/Self-Supervised Deep Learning Models

Introduction : This tutorial introduces Deep Internal Learning (DIL), a family of methods that train deep networks without requiring prior examples. Unlike supervised learning, which depends on large annotated datasets, DIL leverages the internal structure and self-similarity within the image itself. We begin with the core motivation and taxonomy of DIL approaches, covering both learning-from-scratch and test-time adaptation techniques. The tutorial then explores applications in image restoration (denoising, inpainting, superresolution), followed by style transfer and image decomposition frameworks. Next, we discuss Single-Image GANs for object synthesis and GAN inversion for restoration. The tutorial concludes with a discussion on challenges and opportunities for future research.

Outline and Duration :

• Introduction (20 Mins): Motivation for DIL; contrast with supervised learning; taxonomy: learning from scratch vs test-time adaptation; typical application domains such as image restoration and synthesis.
• Image Restoration by Reconstruction (30 mins): Denoising, Inpainting, Superresolution; optimizing untrained networks; early stopping; example frameworks and architectures.
• Image Decomposition (30 mins): Layer separation from multiple Deep Image Priors for segmentation, dehazing, transparency separation, and watermark removal.
• Style Transfer (30 mins): Style transfer introduction; text-guided stylization via CLIP; multimodal inputs
• Single Image GANs for Synthesis (30 mins): Learning generative models from a single image; object synthesis and retargeting; visual diversity from a single instance.
• GAN Inversion for Restoration (30 mins): Using pretrained StyleGAN for image restoration; working with latent codes.
• Future Directions (10 mins): Summary of progress in DIL; open challenges, interactive discussion, and Q&A.

Instructor :

Dr. Indra Deep Mastan is an Assistant Professor at IIT (BHU), Varanasi. He received his Ph.D. from IIT Gandhinagar under the supervision of Prof. Shanmuganathan Raman in 2021. He completed his Master’s from IISc Bangalore and Bachelor’s from MNIT Jaipur. His research primarily focuses on deep learning for computer vision, including image restoration, enhancement, style-transfer, and unsupervised and self-supervised learning. He has also been actively involved in delivering invited talks and workshops at prestigious institutions and conferences. He is currently working on research projects on digital manuscript restoration and AI-based rice grain quality analysis. email: indra.cse@iitbhu.ac.in