CS654: Digital Image Analysis Lecture 22: Image Restoration
Recap of Lecture 21 Image restoration in presence of only noise Image restoration in presence of only degradation Observation, experimentation and mathematical modeling Motion blur Restoration by inverse filtering
Outline of Lecture 22 Inverse filtering and its problems Pseudo Inverse filtering Constrained image restoration problem
Inverse filtering (error minimization) 2D Discrete Domain Representation Neglecting the noise component Approximate least square error: Unconstrained error minimization
Inverse filtering (error minimization) Equating to zero In frequency domain It doesn't perform well when used on noisy images.
Pseudo inverse filtering Equation of inverse filter in frequency domain Spectrum of the PSF Simulated impulseImpulse response Pseudo inverse filter
SVD approach to Pseudo-Inverse Image restoration model
Constrained image restoration SmoothnessRestoration
Minimization of error Equating to zero Emphasize restoration Emphasize smoothness
Constrained Restoration By applying Fourier transform matrices to both sides Constrained restoration results with Q = Laplacian and different γ values
Wiener Filters Assume: noise is zero mean and uncorrelated with the image
Weiner Filter Product of a complex quantity with its conjugate is equal to the magnitude of the complex quantity squared
Weiner filter
Weiner Filter Approximation of Weiner filter
Example Input image Full inverseRadially limitedWeiner Filter
Example Noise + Motion InverseWeiner Noise
Conclusion We considered several algebraic approaches to image restoration Constrained restoration imposes smoothness constraints and does well when noise is present Wiener filters model the noise/signal ratio to obtain a minimum mean square error restoration image