| PhD Viva

Name of the Speaker: Mr. Pawan Prasad (EE19D005)
Guide: Dr. Kaushik Mitra
Online meeting link: https://meet.google.com/jht-rmyt-rfj
Date/Time: 20th December 2024 (Friday), 4:30 PM
Title: Fast and Efficient Methods for Image Reflection Removal

Abstract :

The images that are captured in the presence of obstruction such as a glass window create undesirable reflection artifacts. The removal of such undesirable reflections has been an actively researched topic in the field of computer vision. Despite several recent advances, the state of the art methods for reflection removal suffer from following limitations. (a) Computational Complexity: existing methods mostly focus on restoration quality but compromise on processing speed and memory constraints. (b) Applicability to High Resolution Images: current methods use a fixed network architecture with a static receptive field. Moreover, these methods train their models on low resolution images resulting in low generalizability and scalability to high resolution images. (c) Low Transmitted Reflections: existing methods are unable to handle this category of reflections where undesirable reflections occlude the desirable background regions. (d) Severely ill-posed: Single image reflection removal is a severely ill-posed problem. With only a single viewpoint perspective, there is lack of information to separate the desirable background from the undesirable reflections. This ill-posed nature necessitates the use of additional constraints or prior knowledge to guide algorithms and make it more tractable. We firstly present deep learning architectures for removing different categories of reflections from a single high resolution image at low computational complexity. Our models are incredibly light weight, consume very less memory and are capable of running in real time on resource limited devices such as smartphones. Secondly, we demonstrate how discriminating cues are exploited from additional modality of adding a burst of images to further aid reflection removal. Further, we propose a reference based approach for reflection removal where a multi-stage architecture is deployed to first extract novel cues from a given reference image which acts as a guide to efficiently remove reflections. We also extend this idea to build a large semantic feature dictionary using which the reflection removal capabilities are further enhanced.