公开(公告)号：US20180061018A1

公开(公告)日：2018-03-01

申请号：US15630688

申请日：2017-06-22

Applicant: KOREA INSTITUTE OF SCIENCE AND TECHNOLOGY

Inventor： Byeongjoo AHN ,  Ig Jae KIM ,  Junghyun CHO

IPC: G06T5/00 ,  G06T7/579 ,  H04N13/00

Abstract: A method of multi-view deblurring for 3-dimensional (3D) shape reconstruction includes: receiving images captured by multiple synchronized cameras at multiple viewpoints; performing iteratively estimation of depth map, latent image, and 3D motion at each viewpoint for the received images; determining whether image deblurring at each viewpoint is completed; and performing 3D reconstruction based on final depth maps and latent images at each viewpoint. Accordingly, it is possible to achieve accurate deblurring and 3D reconstruction even from any motion blurred images.

公开(公告)号：US20170011494A1

公开(公告)日：2017-01-12

申请号：US15191040

申请日：2016-06-23

Applicant: Korea Institute of Science and Technology

Inventor： Byeongjoo AHN ,  Ig Jae KIM ,  Junghyun CHO

IPC: G06T5/00 ,  H04N5/217 ,  G06T5/50 ,  H04N5/14

CPC classification number: G06T5/003 ,  G06T2207/10016 ,  G06T2207/20201

Abstract: A video deblurring method based on a layered blur model includes estimating a latent image, an object motion and a mask for each layer in each frame using images consisting of a combination of layers during an exposure time of a camera when receiving a blurred video frame, applying the estimated latent image, object motion and mask for each layer in each frame to the layered blur model to generate a blurry frame, comparing the generated blurry frame and the received blurred video frame, and outputting a final latent image based on the estimated object motion and mask for each layer in each frame, when the generated blurry frame and the received blurred video frame match. Accordingly, by modeling a blurred image as an overlap of images consisting of a combination of foreground and background during exposure, more accurate deblurring results at object boundaries can be obtained.

Abstract translation: 基于分层模糊模型的视频去模糊方法包括：在接收到模糊视频帧时，在相机的曝光时间期间，使用由层的组合构成的图像来估计每个帧中的每个层的潜像，物体运动和掩模， 将每个帧中的每个层的估计潜像，对象运动和掩模应用于分层模糊模型以生成模糊帧，比较生成的模糊帧和接收到的模糊视频帧，并基于估计对象输出最终潜像 当生成的模糊帧和接收的模糊视频帧匹配时，每个帧中的每个层的运动和掩码。 因此，通过将模糊图像建模为在曝光期间由前景和背景的组合组成的图像的重叠，可以获得对象边界处的更准确的去模糊结果。