We assume a standard acquisition model. The original image undergoes two degradations during the measurement: blurring and corruption by noise.

The blur may be caused by several external effects, such as, atmospheric turbulence, camera lens, relative camera-scene motion, etc. Yet we will assume that we can model them as convolution with an unknown point spread function (PSF).

We have P blurred and noisy images z_k and the goal of blind deconvolution is to estimate the original image u without any knowledge of blur kernels h_k.

Multichannel blind deconvolution approaches required perfectly registered images. We have proofed that the inaccurate registration of images can be alleviated by properly overestimating a blur size adn we have proposed a novel blind deconvolution algorithm that can deal with the overestimated blur size. Alternating minimization (AM) scheme is used to find a maximum a posteriori (MAP) estimator.

Code to download:

Matlab code with GUI.

Examples:

Real data sets to download: (input images and our PSF and image reconstruction)

dataset1, dataset2, dataset3, dataset4

Blind deconvolution of long-exposure photos

The example below shows restoration of two real photos blurred by camera motion. On the left are two input images and on the right is the MAP estimator and reconstructed PSFs. The size of PSFs was set to (15x15).

Blind deconvolution of sunspot images

A sequence of short-exposure images of a sunspot was acquired with a terrestrial telescope. The images were taken shortly one after another and hence the image alignment was more or less assured and no registration was necessary. Atmospheric turbulence is the main cause of the image degradation but through its temporal fluctuations also provides the necessary co-primeness of channels. The least degraded image from the sequence, which is shown at the bottom right, was selected as a reference image. Two other images of median degradation (top row) were used as the input for the algorithm. The size of blurs was set to (12x12) which we believe to be large enough to contain the original blurs. The estimated original image at the bottom left was obtained after 3 iterations of AM.

Severly blurred images
	Restored image and PSFs	Reference image

Contact person: Filip Šroubek

Publications:

• Šroubek Filip, Flusser Jan: Multichannel blind deconvolution of spatially misaligned images, IEEE Transactions on Image Processing vol.14, 7 (2005), p. 874-883
• Šroubek Filip, Flusser Jan: Multichannel blind iterative image restoration, IEEE Transactions on Image Processing vol.12, 9 (2003), p. 1094-1106
• Šroubek Filip, Šimberová Stanislava, Flusser Jan, Suk Tomáš: Multichannel blind deconvolution of the short-exposure astronomical images, Proceedings of the 15th IAPR International Conference on Pattern Recognition, p. 53-56, Eds: Sanfeliu A., Villanueva J. J., Vanrell M., IEEE Computer Society, (Los Alamitos 2000), IAPR International Conference on Pattern Recognition /15./, (Barcelona, ES, 03.09.2000-07.09.2000)
• Šroubek Filip, Flusser Jan: An overview of multichannel image restoration techniques, Week of Doctoral Students 1999, p. 580-585, Eds: Šafránková J., MATFYZPRESS, (Praha 1999), Annual Conference of Doctoral Students. WDS '99 /8./, (Praha, CZ, 22.06.1999-25.06.1999)