À la Une

BM3D-based denoising and super-resolution

Presenter: Karen Egiazarian (Eguiazarian), Prof. , Technical University of Technology, Tampere, and CEO, Noiseless Imaging Oy (Finland)

karen00_560x420.jpg

Abstract:

We first demonstrate the BM3D denoising filter on an iPhone 6, where we utilize parallel-multicore processing together with NEON SIMD instructions on CPU.

We also developed efficient OpenCL implementations of our BM3D for mobile, leveraging both CPU and GPU and providing very high performance (e.g., 16-Mpix Bayer raw data in 130 ms on an ultrabook with Intel i7-4600U processor).

Next, we discuss the filterís application to super-resolution (SR), i.e. for reconstructing high-resolution (HR) image(s) from a set of low-resolution (LR) input image(s).

We show that BM3D-based single-image SR (also known as image upsampling) outperforms the state-of-the-art methods based on external dictionary learning and deep convolutional neural networks (CNN). Finally, we demonstrate BM3D-based video SR (with increase of both spatial and temporal rate) as well as cross-modal video SR.

In one application, the scene is captured by a LR color camera with Bayer sensor and by a second camera with panchromatic sensor; the goal is to obtain a HR full-color video by super-resolving the luminance from the second camera and combining it with the chrominance from the first camera.

In another application, the scene is captured by a LR thermal imager and by a standard color camera; here the goal is to super-resolve the thermal video with the aid of the color video. In either case, the cameras do not need to be bore-sighted and can be located several inches apart.

Short Bio:

Karen O. Egiazarian (Eguiazarian) (Senior Member IEEE, 1996) received M.Sc. in mathematics from Yerevan State University, Armenia, in 1981, the Ph.D. degree in physics and mathematics from Moscow State University, Russia, in 1986, and a D.Tech from Tampere University of Technology, Finland, in 1994.

In 2015 he has received the Honorary Doctoral degree from Don State Technical University (Rostov-Don, Russia).

Dr. Egiazarian is a co-founder and CEO of Noiseless Imaging Oy (Ltd), Tampere University of Technology spin-off company.

He is a Professor at Signal Processing Department, Tampere University of Technology, Tampere, Finland, leading ‘Computational imaging’ group and a Docent in the Department of Information Technology, University of Jyväskyla, Finland. His main interests are in the field of computational imaging, compressed sensing, efficient signal processing algorithms, image/video restoration and compression. Dr. Egiazarian has published about 650 refereed journal and conference articles, books and patents in these fields. He is an Editor-in-Chief of Journal of Electronic Imaging (SPIE), Associate Editor of IEEE Transactions of Image Processing, and Member of the DSP Technical Committee of the IEEE Circuits and Systems Society.

Date: Wednesday 23th November 2016, 14h00

Location: Auditorium ground floor, Battelle building A

24 octobre 2016
  À la Une