US 8,428,390 B2 · Generating sharp images, panoramas, and videos from motion-blurred videos

Abstract

A “Blur Remover” provides various techniques for constructing deblurred images from a sequence of motion-blurred images such as a video sequence of a scene. Significantly, this deblurring is accomplished without requiring specialized side information or camera setups. In fact, the Blur Remover receives sequential images, such as a typical video stream captured using conventional digital video capture devices, and directly processes those images to generate or construct deblurred images for use in a variety of applications. No other input beyond the video stream is required for a variety of the embodiments enabled by the Blur Remover. More specifically, the Blur Remover uses joint global motion estimation and multi-frame deblurring with optional automatic video “duty cycle” estimation to construct deblurred images from video sequences for use in a variety of applications. Further, the automatically estimated video duty cycle is also separately usable in a variety of applications.

Claims20

Dependent claims

A method for deblurring images in an image sequence, comprising steps for:

receiving an input image sequence comprising a plurality of motion-blurred images;
evaluating the image sequence to initialize motion estimates in each image by computing optical flow for each image;
computing a current blur kernel from the motion estimates;
generating a current set of deblurred images from the image sequence using the current blur kernel;
until convergence of the current blur kernel is achieved, iteratively repeating steps for:
using the current set of deblurred images to minimize a multi-image deconvolution energy function to produce refined motion estimates for each image of the input image sequence;
computing a new current blur kernel from the refined motion estimates;
generating a new current set of deblurred images from the image sequence using the new current blur kernel;
and
checking for convergence of the new current blur kernel relative to the immediately preceding blur kernel.

7 more elements

A computer-readable memory having computer executable instructions stored therein for deblurring images, said instructions comprising:

evaluating a sequence of motion-blurred images to initialize motion estimates in each image by computing optical flow for each image;
computing a current blur kernel from the motion estimates and an initial camera duty cycle value;
generating a current set of deblurred images from the image sequence using the current blur kernel;
iteratively computing a new current blur kernel for use in deblurring the image sequence until convergence of the current blur kernel is achieved, comprising:
using the current set of deblurred images to jointly minimize a multi-image deconvolution energy function to produce refined motion estimates for each image of the input image sequence in combination with a refined camera duty cycle value;
computing a new current blur kernel from the refined motion estimates and the refined camera duty cycle value;
generating a new current set of deblurred images from the image sequence using the new current blur kernel;
and
checking for convergence of the new current blur kernel relative to the immediately preceding blur kernel.

6 more elements

Dependencies are inferred from the claim text.

Generating sharp images, panoramas, and videos from motion-blurred videos

Life of this patent

Generating sharp images, panoramas, and videos from motion-blurred videos

Life of this patent

Abstract

Claims20