Blind Deconvolution

The goal of the general blind deconvolution problem is to recover two convolved signals, f and h, when only a noisy version of their convolution is available along with some or no partial information about either signal. Some important characteristics of the problem of blind deconvolution is listed below: 1. 2. 3. 4. 5. The true image and the point-spread function must be irreducible for an unique solution to the problem. A signal is reducible if it can be expressed as the convolution of two other signals (neither of which is the two-dimensional delta function). If either the true image or the point-spread function is reducible, then there exists more than one way to perform the deconvolution.

2. Without the use of appropriate a priori information, only a scaled, shifted version of the original image can be obtained through blind deconvolution. 3. In the presence of additive noise, the exact blind deconvolution of the observed image g(x,y) is impossible and only an approximate solution can be found. The point spread function (PSF) describes the response of an imaging system to a point source or point object . A more general term for the PSF is a system's impulse response, the PSF being the impulse response of a focused optical system. The PSF in many contexts can be thought of as the extended blob in an image that represents an unresolved object

Iterative Blind Deconvolution

Starting with a random initial guess for the true image, the algorithm alternates between the image and Fourier dom using known constraints in each domain. The image domain constraints are simply those imposed by the nonnega requirement and the known finite support information. Negative-valued pixels and nonzero pixels outside the region support are set to zero. The Fourier domain constraint is a consequence of the fact that if the noise level is low, the product of the Fourier transforms of the true image and the point-spread function should be approximately equal to Fourier transform of the observed image. Using this fact, we can use inverse filtering to estimate the point-spread f (image) using the Fourier transforms of the degraded observation and the estimate of the true image (point-spread

The MATLAB code for the same is as shown below: % Read the image >>f=imread('G:college.jpg'); % Showing the actual picture >> imshow(f); %Simulate a real-life image that could be blurred (e.g., due to camera %motion or lack of focus). The example simulates the blur by convolving a %Gaussian filter with the true image (using imfilter). The Gaussian filter %then represents a point-spread function, PSF. >> PSF = fspecial('gaussian',7,10); Blurred = imfilter(f,PSF,'symmetric','conv'); >> imshow(Blurred); >> title('Blurred Image'); %To illustrate the importance of knowing the size of the true PSF, this example %performs three restorations. Each time the PSF reconstruction starts from a %uniform array--an array of ones. %The first restoration, J1 and P1, uses an undersized array, UNDERPSF, for an initial

%guess of the PSF. The size of the UNDERPSF array is 4 pixels shorter in each dimension than the true PSF >> UNDERPSF = ones(size(PSF)-4); [J1 P1] = deconvblind(Blurred,UNDERPSF); imshow(J1); title('Deblurring with Undersized PSF'); %The second restoration, J2 and P2, uses an array of ones, OVERPSF, for an %initial PSF that is 4 pixels longer in each dimension than the true PSF. >> OVERPSF = padarray(UNDERPSF,[4 4],'replicate','both'); [J2 P2] = deconvblind(Blurred,OVERPSF); imshow(J2); title('Deblurring with Oversized PSF'); %The third restoration, J3 and P3, uses an array of ones, INITPSF, for an %initial PSF that is exactly of the same size as the true PSF. >> INITPSF = padarray(UNDERPSF,[2 2],'replicate','both'); [J3 P3] = deconvblind(Blurred,INITPSF); imshow(J3); title('Deblurring with INITPSF'); %All three restorations also produce a PSF. The following pictures show how %the analysis of the reconstructed PSF might help in guessing the right size for the initial PSF. In the true PSF, %a Gaussian filter, the maximum values are at the center (white) and diminish at the borders (black). >> imshow(J1); imshow(PSF,[],'InitialMagnification','fit'); title('True PSF'); >> imshow(P1,[],'InitialMagnification','fit'); title('Undersized PSF'); >> imshow(J1); imshow(PSF,[],'InitialMagnification','fit'); title('True PSF'); >> imshow(P2,[],'InitialMagnification','fit'); title('Oversized PSF'); >> imshow(P3,[],'InitialMagnification','fit'); title('Reconstructed True PSF'); %The PSF reconstructed in the first restoration, P1, obviously does not fit into the constrained size. %It has a strong signal variation at the borders. The corresponding image, J1, does not show any %improved clarity vs. the blurred image, Blurred. %The PSF reconstructed in the second restoration, P2, becomes very smooth at the edges. % This implies that the restoration can handle a PSF of a smaller size. The corresponding image, J2, %shows some deblurring but it is strongly corrupted by the ringing. %Finally, the PSF reconstructed in the third restoration, P3, is somewhat intermediate between P1 and P2. %The array, P3, resembles the true PSF very well. The corresponding image, J3, shows significant improvement; %however it is still corrupted by the ringing.

Original Image