image deblurring

objective: compensate the motion blur caused by the relative motion between the camera and the scene

accurate spatially-variant motion deblurring using camera motion tracking and scene depth

objective: in reality, the motion experienced by each pixel in a blurry image is not same (or spatially-invariant). we want to estimate the blur kernels by reconstructing the camera motion during the exposure time from the inertial measurement unit data and depth sensor (Microsoft Kinect)


System configuration: Our system is composed of the DSLR Camera, IMU (inertial measurement unit) for the camera motion tracking, the depth sensor (Microsoft Kinect) and the control board (Beagleboard-xm).


Algorithm overview: The camera motion is reconstructed by compensating the gravity and drift components from the IMU data. The estimated camera motion is combined with the depth map of the scene to estimate the trajectory of each pixel during the exposure time (or spatially-variant blur kernel).


Estimated results: (a)original blurry image (3 different areas are selected to show estimated blur kernels and latent images) Our results are provided in (c),(h),and (m). The estimated blur kernels for center of the dotted rectangles in (a) are shown besides of the boxes of (a).

Hyeoungho Bae, Charless C. Fowlkes, Pai H. Chou, Accurate Motion Deblurring using Camera Motion Tracking and Scene Depth, WACV, Clear Beach, Florida, (Jan. 2013).

performance: Our algorithm can estimate spatially-variant blur kernel with minimal limit on the degree of freedom in the camera motion or limit on the scene depth (we have all the depth data for every pixel in the scene and our camera motion has 6-degree of freedom).


author = {Hyeoungho Bae and Charless C. Fowlkes and Pai H. Chou},
title = {Accurate Motion Deblurring using Camera Motion Tracking and Scene Depth},
booktitle = {in WACV},
year = 2013

poster, paper

patch mosaic for fast motion deblurring

objective: use a mosaic image of informative image patches found in a blurry image


(a) Conventional MAP_i,k blind motion deblurring algorithm


(b) Patch Mosaic based motion deblurring algorithm

Algorithm overview: Patch Mosaic Algorithm (yello box) can be plugged into several blind motion deblurring algorithm without significant modification.


(a) The patch mosaic and estimated blur kernel during the iteration, (b) The original blurry image, and (c) The estimated latent image using our Patch Mosaic based algorithm

Hyeoungho Bae, Charless C. Fowlkes, Pai H. Chou, Patch Mosaic Algorithm for Fast Motion Deblurring, ACCV (Asian Conference on Computer Vision), Daejeon, Korea, (Nov. 2012).

performance: our algorithm reduces the image data used for the blur kernel estimation down to 30%. The time consumption for estimating latent image (2256 by 1504 color image) is 14.9sec for Matlab script, which is 24% faster than our image patch based fast motion deblurring algorithm.

executable file (If you need matlab MCRinstaller.exe, send me e-mail.)


author = {Hyeoungho Bae and Charless C. Fowlkes and Pai H. Chou},
title = {Patch Mosaic for Fast Motion Deblurring},
booktitle = {in ACCV},
year = 2012

poster, paper

image patch based fast motion deblurring

objective: reduce the computation time of motion deblurring to expand the application area to portable imaging devices (e.g. DSLR, smart phones, ...)

performance: our Matlab script algorithm is 4 to 100 times faster than the c\c++ binaries of state-of-the-art motion deblurring algorithms with similar or better estimation accuracy.

overview of fast motion deblur algorithm

fast motion deblur results

Hyeoungho Bae, Charless C. Fowlkes, Pai H. Chou, Fast Motion Deblurring, ICCP12 Poster, Seattle, WA, 2012.

poster_button Poster file