Gyro-Based Multi-Image Deconvolution for Removing Handshake Blur

CVPR, June 2014

Deblurring results from our system. Eight images are captured from a single burst at 5M-pixel resolution, while gyroscope data is recorded simultaneously. Each column in (a) and (b) shows two insets of an input image with their blur kernels shown at 2x size. Note the difference in shape of blur in two insets, which demonstrates spatially-varying blur. Input images are jointly deblurred with two different priors. The deblurred images in (d) and (e) removed handshake blur more effectively than in (c), where the input images are aligned and averaged.

Image deblurring to remove blur caused by camera shake has been intensively studied. Nevertheless, most methods are brittle and computationally expensive. In this paper we analyze multi-image approaches, which capture and combine multiple frames in order to make deblurring more robust and tractable. In particular, we compare the performance of two approaches: align-and-average and multi-image deconvolution. Our deconvolution is non-blind, using a blur model obtained from real camera motion as measured by a gyroscope. We show that in most situations such deconvolution outperforms align-and-average. We also show, perhaps surprisingly, that deconvolution does not benefit from increasing exposure time beyond a certain threshold. To demonstrate the effectiveness and efficiency of our method, we apply it to still-resolution imagery of natural scenes captured using a mobile camera with flexible camera control and an attached gyroscope.

Handling Moving Objects and Over-Exposed Regions
in Non-Blind Multi-Image Deconvolution

Stanford Computer Graphics Laboratory Technical Report 2014-01

Output images after combining the results of multi-image deconvolution and patch-based multi-image denoising. Eight images are captured from a single burst at 5M-pixel resolution, while gyroscope data is recorded at the same time. (a) A reference image is denoised by a patch-based method. (b) Our multi-image deconvolution is applied. Then, (a) and (b) are blended to obtain the final output images shown in (c). Close-ups, shown as yellow boxes in (a)-(c), are shown in (e)-(h). Note that the reference images in (e) are noisy and contain some amount of blur, while noise is reduced in (f). The deconvolved images in (g) recovered sharp details in static regions (top row), but moving objects suffer from motion blur (middle row) and highlights caused artifacts (bottom row). The blended results in (h) effectively hide artifacts in (g) while preserving sharpness in the background. Also note that outliers, i.e., moving objects and over-exposed regions, are well detected with our method as shown in (d).