1. Image Modeling & Segmentation Aly Farag and Asem Ali Lecture #2

2. Intensity Model “Density Estimation”

3. 3 Intensity Models  The histogram of the whole image will represent the rate of occurrence for all the classes in the given empirical density  Each class can be described by the histogram of the occurrences of the gray levels within that class  Intensity models describe the statistical characteristics of each class in the given image. marginal density  T he objective of the intensity model is to estimate the marginal density for each class from the mixed normalized histogram of the occurrences of the gray levels

4. 4  Density estimation can be studied, under two primary umbrellas:  Parametric methods and  Nonparametric methods. Intensity Models  Nonparametric methods take a strong stance of letting the data (e.g., pixels’ gray levels) represent themselves.  One of the core methods that nonparametric density estimation approaches based on is the k-nearest neighbors (k-NN) method. These approaches calculate the probability of a sample by combining the memorized responses for the k nearest neighbors of this sample in the training data.  Nonparametric methods achieve good estimation for any input distribution as more data are observed.  Flexible: they can fit almost any data well. No prior knowledge is required.  However, they apparently often have a high computational cost and have many parameters that need to be tuned.

5. 5 Nonparametric methods 1  Based on the fact that the probability P that a given sample x falls within a region (window) of R is given by  This integral can be approximated either by the product of the value of p(x) with the area/volume of the region, or by the ratio of number of samples fall within the region:  if we observe a large number, n, of fish and count those whose length fall within the range defined by R, then k/n can be used as n estimate of P as n→∞ 1 Chapter 4, Duda, R., Hart, P., Stork, D., 2001. Pattern Classification, 2nd edR John Wiley & Sons.

6. 6 Nonparametric methods  In order to make sure that we get a good estimate of p(x) at each point, we have to have lots of data points (instances) for any given R (or volume V ). This can be done in two ways  We can fix V and take more and more samples in this volume. Then k/n→P, however, we then estimate only an average of of p(x), not the p(x) itself, because P can change in any region of nonzero volume  Alternatively, we can fix n and make V →0, so that p(x) is constant in that region. However, in practice we have a finite number of training data, so as V →0, V will be so small that it will eventually contain no samples: k =0 p(x) = 0, a useless result!  Therefore, V →0 is not feasible and there will always be some variance in k/n and hence some averaging in p(x) within the finite non-zero volume V. A compromise need to be found for V so that  It will be large enough to contain sufficient number of samples  It will be small enough to justify our assumption of p(x) be constant within the chosen volume/region.

7. 7 Nonparametric methods To make sure that k/n is a good estimate of P, and consequently, p n (x*) is a good estimate p(x*) the following need to be satisfied:

8. 8 Nonparametric methods  There are two ways to ensure these conditions:  Shrink an initial volume V n as a function of n, e.g.,. Then, as n increases so does k, which can be determined from the training data. Parzen Windows (PW) density estimation  Specify k n as a function of n, e.g., V n grows until it encloses k n samples. Then V n can be determined from the training data. K- Nearest Neighbor (KNN)  It can be shown that as n→∞, both KNN and PW approach the true density p(x*), provided that V n shrinks and k n grows proportionately with n.

9. 9 Parzen Windows  The number of samples falling into the specified region is obtained by the help of a windowing function, hence the name Parzen windows. We first assume that R is a d -dimensional hypercube of each side h, whose volume is then V=(h) d Then define a window function φ(u), called a kernel function, to count the number of samples k that fall into R

10. 10 Parzen Windows Example:Given this image D = {1,2,5,1,1,2,3,5,5,1.5,1.5} For h=1, compute p(2.5) d = V = n = K = p(2.5) =

11. 11 Parzen Windows  Now consider ϕ (.) as a general function, typically a smooth and continuous function, instead of a hypercube. The general expression of p(x) remains unchanged  Then p(x) is an interpolation of ϕ (.) s, where each ϕ (.) measures how far a given x i is from x  In practice, x i are the training data points and we estimate p(x) by interpolating the contributions of each sample data point x i based on its distance from x, the point at which we want to estimate the density. The kernel function ϕ (.) provides the numerical value of this distance  If ϕ (.) is itself a distribution, then p(x) will converge to p(x) as n increases. A typical choice for ϕ (.) is the Gaussian  The density p(x) is then estimated simply by a superposition of Gaussians, where each Gaussian is centered at the training data instances. The parameter h is then the variance of the Gaussian

12. 12 Assume you have n samples drawn from normal distribution N(0,1). Use PW with Gaussian kernel to estimate this distribution. Try different window widths and numbers of samples. i.e., Try to generate a similar figure. Image Modeling Homework #2 due Sept. 1 st

13. 13 Classification using Parzen Windows  In classifiers based on Parzen-window estimation, we estimate the densities for each class and classify a test point by the label corresponding to the maximum posterior. 10,c 1 20,c 1 15,c 1 23 3,c 1 5 50,c 2 60,c 2 43 30,c 2 29 40,c 2 Example: