1. Lecture 10 Chapter 5: Image Restoration

2. Image restoration Image restoration is the process of recovering the original scene from the observed scene which is degraded. Different form enhancement- aim of enhancement techniques make images visually appealing, whereas restoration essentially inverts the degradation- more objective.

3. Image restoration Has intersections with signal processing, estimation theory, inverse problems, linear algebra. As with most of the topics in Image Processing, restoration was developed to be used in astronomy, and is also used for restoring old films and pictures

4. Degradation model Image degradation model: – g : Observed degraded image – h: Impulse response of the system – n: noise Noise is assumed to be uncorrelated with pixel values and does not depend on spatial coordinates

5. Noise models Noise is characterized by its probability density function. Some of which are: Gaussian, Rayleigh, Uniform, Exponential, etc. By far Gaussian is the most popular model, because: – It is present widely in practice – Mathematical ease

6. Noise models

7. Noise on images

8. Identifying noise model If the imaging system is available we can capture images of a uniform gray value object, example a black board uniformly illuminated. If images from the system are given, look for part of images which are uniform.

9. Identifying noise Observing the histogram we can figure out which model the noise comes from.

10. Estimation of noise parameters Noise is characterized by the pdf, which in turn depends on certain parameters. For example the Gaussian model depends only on the mean and variance and can easily be computed from the histogram. where p(zi) is the normalized histogram values.

11. Restoration in presence of noise Degradation model: g = f + n Restoration = Enhancement –Mean filtering, Median filtering, Notch filters. Some new ones: – Geometric mean filter, Harmonic mean filters, Midpoint filter, etc.

12. Geometric mean filtering Comparable to arithmetic mean filter, but introduces less blurring

13. Contra-harmonic mean filtering Q > 0 – for pepper noise Q < 0 – for salt noise

14. Alpha-trimmed mean filter Let Sxy be a neighborhood of size M x N, of a point (x,y) in the image I. Given a number d, delete d/2 largest and d/2 smallest gray-level values in Sxy. Let the remaining gray-level values be denoted as I d. Take the arithmetic mean of rest of the gray-level values: – –Since it is a combination of median and mean filtering, it can be used in cases where there are multiple kinds of noise, typically Gaussian and Salt & pepper noise.

15. Alpha-trimmed filter output A1 = Image + uniform noise A2 = A1 + Salt-pepper noise Arithmetic mean outputGeometric mean output Median filter output Alpha-trimmed mean filter output, d = 5

16. Adaptive filters for handling noise Adapts to some statistical measures on a local neighborhood, usually mean and variance: – g(i,j): the noisy image gray-level value. – : noise variance in the image. – : local mean in neighborhood. – : local variance in neighborhood. could be estimated from an uniform area in the given image.

17. Adaptive local noise reduction filter If noise variance is zero  Indicates no noise  return the observed image. If local variance is high compared to the noise variance  presence of edge or a sharp feature  return a value close to observed grey-value. If two variances are equal  presence of noise  return the arithmetic mean

18. Adaptive local noise reduction filter Filter equation: – How to handle, what does it mean? When, let in the filter equation.

19. Adaptive filter output

20. Adaptive noise filters Adaptive median filter: –Let S xy be a neighborhood of pixel (x,y). –Let S max be the maximum allowed size of neighborhood. –Z min be the minimum gray value in S xy –Z max be the minimum gray value in S xy –Z xy be the gray level at pixel (x,y) –Z med be the median of gray values in S xy.

21. Adaptive median filtering Level A: A1 = z med – z min, A2 = z med – z max If A1 > 0 and A2 < 0, Go to level B Else increase window size If window size <= S max, repeat level A Else output z xy Level B: B1 = z xy – z min, B2 = z xy – z max If B1 > 0 and B2 < 0, output z xy Else output z med

22. Experiments

23. Frequency domain noise filtering Bandpass, Band-reject, Notch filters

24. Frequency domain noise filtering A B -Fourier spectra of AC - Filtered spectra D – Inverse Fourier transform of C Inverse Fourier transform of (B-C)