News & Events

Talk by Dr. Kuldeep Purohit

Date/Time: 
Friday, June 4, 2021 - 19:00
Venue: 
Online
Speaker: 
Dr. Kuldeep Purohit
Affiliation: 
Michigan State University
Title: 

Spatially-Adaptive Image Restoration using Distortion-Guided Networks

Images often undergo degradation during the data acquisition process, especially under non-ideal imaging conditions. Image restoration, which refers to recovering a clean image from the degraded observation, is vital to not only improve the aesthetic quality of the image, but also the performance of downstream tasks such as object detection, semantic segmentation, tracking, and many more. In this talk, we will primarily discuss a new approach suitable for removing practically occurring artifacts such as rain-streaks, haze, raindrops, motion blur, and shadows, which are image-specific and spatially-varying in nature. Prior approaches are typically degradation type-specific and employ the same processing across different images and different pixels within. However, we hypothesize that such spatially rigid processing is suboptimal for simultaneously restoring the degraded pixels as well as reconstructing the clean regions of the image. To overcome this limitation, we propose a network design that harnesses distortion-localization information and dynamically adjusts computation to difficult regions in the image. It comprises two components, (1) a localization network that identifies degraded pixels, and (2) a restoration network that exploits knowledge from the localization network in filter and feature domain to selectively and adaptively restore degraded pixels. We exploit the knowledge gained by the localization network to guide the restoration network’s training using attentive knowledge distillation. Our key idea is to exploit the non-uniformity of heavy degradations in spatial-domain and suitably embed this knowledge within distortion-guided modules that perform sparse versions of normalization, convolution, and attention. Since our design is agnostic to the physical formation model and generalizes across several types of spatially-varying degradations and shows multi-task learning capability. Extensive qualitative and quantitative comparisons with prior art demonstrate that our degradation-agnostic network design offers significant performance gains over state-of-the-art degradation-specific architectures. Next, we will briefly discuss the removal of degradations caused due to camera/scene motion and/or poor sensor resolution and investigate the problems of motion deblurring, temporal and spatial super-resolution from a single blurred image. Finally, we shall discuss the current challenges in the image restoration area and potential future directions.

Corporate Site - This is a contributing Drupal Theme
Design by WeebPal.