| PhD Seminar


Name of the Speaker: Mr. Pawan Prasad (EE19D005)
Guide: Dr. Kaushik Mitra
Venue: ESB-234 (Malaviya Hall)
Online meeting link: meet.google.com/qyb-xtsg-ega
Date/Time: 23rd August 2024 (Friday), 3:00 PM
Title: Fast and Efficient Methods to Image Reflection Removal

Abstract :

Images captured in front of a glass obstruction often suffer from degradation due to the presence of reflections. The existing multi-image based reflection removal methods require input images to be captured at sufficiently different view points with large baselines. A more convenient way for the user is to capture a burst of images using a hand-held device such as a smartphone. Our method exploits the cues that rely on subtle handshakes created during the capture process to efficiently remove reflections. We extract these novel cues called as reflection motion aggregation (RMA) cues to guide the reflection removal in a two stage manner. 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. This cue is extracted by matching neural features between the input image corrupted by reflections and the given reflection free reference image. We also extend this idea to build a large database of easily source-able reflection free images that are publicly available. The proposed method initially retrieves a small subset of semantically similar images from this database followed by a novel reflection aware feature extractor (RAFE) to create a semantic feature dictionary (SFD). The SFD contains features that are mostly invariant to the presence of reflections. These SFD features are later used to guide the reflection removal process.