Classification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University

Size: px

Start display at page:

Download "Classification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University"

Diane Lang
6 years ago
Views:

1 Classification Vladimir Curic Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University

2 Outline An overview on classification Basics of classification How to choose appropriate features (feature set) How to perform classification Classifiers

3 Where are we right now? Figure: Classification in image processing.

4 What is intelligence? Ability to separate relevant information Ability to learn from examples and to generalize knowledge so that it can be used in other situations Ability to draw conclusions from incomplete information

5 What is classification? Classification is a process in which individual items (objects/patterns/image regions/pixels) are grouped based on the similarity between the item and the description of the group Figure: Classification process.

6 What is classification? Distinguish between different types of objects Determine which parts of the image belong to the object of interest We would like to create an intelligent system that can draw conclusions from our image data Often used as the last stage in automated image analysis task

7 Classification example 1 Figure: Classification of leaves.

8 Classification example 2 Figure: Histological bone implant images.

9 Some important concepts Arrangements of descriptors are often called patterns Descriptors are often called features The most common pattern arrangement is a feature vector with n dimensions Patterns are placed in classes of objects which share common properties A collection of N classes are denoted ω 1, ω 2,..., ω N

10 Terminology Object = pattern = point = sample = vector Feature = attribute = measurement Classifier = decision function (boundary) Class Cluster

11 Dataset By measuring the features of many objects we construct dataset Object Perimeter Area Label Apple Apple Apple Pear Pear Pear

12 Data set and a representation in the feature space

13 Features Features are the individual measurable heuristic properties of the phenomena Discriminating (effective) features Independent features Features: area, perimeter texture color...

14 What are good features? Each pattern is represented in terms of n features x = (x 1,..., x n ) The goal is to choose those features that allow pattern vectors belonging to different classes to occupy compact and disjoint regions It is application dependent You might try many, many features, until you find the right ones Often, people compute 100s of features, and put them all in a classifier The classifier will figure out which ones are good This is wrong!!!

15 Peaking phenomenon (Curse of dimensionality) Additional features may actually degrade the performance of a classifier This paradox is called peaking phenomenon (curse of dimensionality) Figure: Curse of dimensionality.

16 Feature set Colour? Area? Perimeter?...

17 Dimensionality reduction Keep the number of features as small as possible Measurements cost Accuracy Simplify pattern representation The resulting classifier will be faster and will use less memory A reduction in the number of features may lead to a loss in the discriminatory power and thereby lower the accuracy

Feature extraction Feature extraction is the process of generating features to be used in the selection and classification tasks Reduce dimensionality by

18 Feature extraction Feature extraction is the process of generating features to be used in the selection and classification tasks Reduce dimensionality by (linear or non-linear) projection of n dimensional vector onto m dimensional vector (m < n) May have not clear physical meaning Figure: Feature extraction.

19 Feature selection Feature selection methods choose features from the original set based on some criteria Reduce dimensionality by selecting subset of original features Have clear physical meaning Figure: Feature selection.

20 Choice of a criterion function The main issues in dimensionality reduction is the choice of a criterion function A most commonly used is the classification error of a feature subset Example: A measure of distance between two distributions f 1 and f 2 Mahalanobis distance: Main assumption is that Gaussian distributions have equal covariance matrix Σ D M (f 1, f 2 ) = (m 1 m 2 ) T Σ 1 (m 1 m 2 ) m 1 is a mean of objects in class 1 m 2 is a mean of objects in class 2

21 Feature selection methods Exhaustive search Best individual features Sequential forward selection Sequential backward selection

22 Exhaustive search Evaluate all ( n m) possible subsets Guaranteed to find the optimal subset Very expensive! Example: If n = 50, m = 5 then there are possible subset (classifications) to evaluate

23 Best individual features Evaluate all n features individually Select the best m n individual features Simple and not likely to lead to an optimal subset Example: Feature 1 is best Feature 2 is best Maybe features 3 and 4 outperform features 1 and 2!

24 Sequential forward selection Selects the best feature and then add one feature at a time Once a feature is retained, it cannot be discarded Computationally fast (for a subset of size 2 examine n 1 possible subsets)

25 Sequential backward selection Starts with all n features and successive delete one feature at a time Once a feature is deleted, it cannot be brought back into the optimal subset Require more computation time than sequential forward selection

26 Search algorithms The presented algorithms are suboptimal When should we use forward selection? When should we use backward selection? Is there any optimal algorithms? Branch and bound (divide and conquer method) Genetic algorithms Simulated annealing

27 Supervised vs. Unsupervised classification Supervised First apply knowledge, then classify Unsupervised First classify, then apply knowledge Figure: Supervised and unsupervised classification.

28 Object-wise and pixel-wise classification Object-wise classification Uses shape information to describe patterns Size, mean intensity, mean color, etc. Pixel-wise classification Uses information from individual pixels Intensity, color, texture, spectral information

29 Object-wise classification Lab 2

30 Object-wise classification Segment the image into regions and label them Extract features for each pattern Train classifier on examples with known class to find discriminant function in the feature space For new examples decide their class using the discriminant function

31 Pixel-wise classification 256*256 patterns (pixels) 3 features (red, green and blue color) Four classes (stamen, leaf, stalk and background)

32 Pixel-wise classification The pattern is a pixel in a non-segmented image Extract (calculate) features for each pattern (pixel) e.g., color, gray-level representation of texture, temporal changes. Train classifier New samples are classified by classifier

33 Train and classify Training Find rules and discriminant function that separate patterns in different classes using known examples Classification Take a new unknown example and put it into the correct class using discriminant function

34 Training data Training data can be obtained from available training samples Classify patterns which are not used during the training stage The performance of classifier depends on the number of available training samples as well as the specific values of the samples Number of training samples should not be too small!

35 Training data Conditional densities are unknown and they have to be learned from available training data Maybe density is known (for example multivariate Gaussian) A common strategy is to replace unknown parameters by the their estimated values Less available information - the difficulty of classification problem increases

36 How to choose appropriate training set?

37 How to choose appropriate training set? Figure: Original image and training image.

38 How to choose appropriate training set? Figure: Original image and training image.

39 Classifiers Once a feature selection finds a proper representation, a classifier can be designed using a number of possible approaches The performance of classifier depends on the interrelationship between sample size, number of features and classifier complexity The choice of a classifier is a difficult problem!!! It is often based on which classifier(s) happen to be available or best known to the user

40 Most commonly used classifiers The user can modify several associated parameters and criterion function There exist some classification problem for which they are the best choice Classifiers (covered in this lecture): Minimal distance classifier Maximal likelihood classifier Nearest neighbour classifier

41 Discriminant functions A discriminant function for a class is a function that will yield larger values than functions for other classes if the pattern belongs to the class d j (x) > d j (x) j = 1, 2,..., N; j i For N pattern classes, we have N discriminant functions The decision boundary between class i and class j is d j (x) d i (x) = 0

42 Decision boundary

43 Decision boundary Figure: Decision boundary.

44 Linear and quadratic classifier Figure: Linear (left) and quadratic (right) classifier.

45 Thresholding - simple classification Classify image into foreground and background Figure: Original image and tresholded image.

46 Box classification Generalized thresholding (multispectral thresholding) Interval for each class and feature All objects with feature vector within the same box belong to the same class

47 Bayesian classifiers Based on a priori knowledge of class probability Cost of errors Minimum distance classifier Maximum likelihood classifier

48 Minimum distance classifier Each class is represented by its mean vector Training is done using the objects (pixels) of known class Mean of the feature vectors for the object within the class is calculated New objects are classified by finding the closest mean vector

49 Minimum distance classifier

50 Limitations of Minimum distance classifier Useful when distance between means is large compared to randomness of each class with respect to its mean Optimal performance when distributions of the classes form spherical shape

51 Maximal likelihood classifier Classify according to the largest probability (taking variance and covariance into consideration) Assume that distribution within each class is Gaussian The distribution in each class can be described by a mean vector and covariance matrix

52 Maximal likelihood classifier For each class from training data compute: Mean vector Covariance matrix Form decision function for each class New objects are classified to class with highest probability

53 Variance and covariance Variance : spread a randomness for class Covariance : influence (dependency) between different features Described by covariance matrix Feature 1 Feature 2 Feature 3 Feature 1 1 1&2 1&3 Feature 2 1&2 2 2&3 Feature 3 1&3 2&3 3

54 Computation of covariance Features : x 1, x 2,..., x n Feature vector for object i: x 1,i, x 2,i,..., x n,i Mean for each class x mean1, x mean2,..., x meanm n Covariance: cov(x i, x j ) = 1 n 1 (x i,k x meani )(x j,k x meanj ) k=1 cov(x 1, x 1 ) cov(x 1, x 2 ) cov(x 1, x 3 ) Σ = cov(x 2, x 1 ) cov(x 2, x 2 ) cov(x 2, x 3 ) cov(x 3, x 1 ) cov(x 3, x 2 ) cov(x 3, x 3 )

55 Density functions Equal variance Different variance Covariance is different from zero

56 Assumptions on covariance matrix 1 Case 1 (Minimal distance) No covariance, equal variance 2 Case 2 (Equal covariance) Same covariance for all classes 3 Case 3 (Uncorrelated) No covariance, different variance 4 Case 4 (General) Different covariances

57 Case 1 Independent features Equal variance Minimal distance classifier

58 Case 2 Equal covariances for all classes

59 Case 3 Independent features - no covariance Different variance for different features

60 Case 4 General Complicated decision boundaries

61 Nearest neighbour Stores all training samples Assigns pattern to majority class among k nearest neighbour

62 k nearest neighbour Assigns pattern to majority class among k nearest neighbour

63 k nearest neighbour Metric dependent Lazy The function is approximated only locally The best choice of k depends on data Sensitive to outliers Larger values of k reduce effects of noise, but make boundaries between the classes less distinct

64 Combine classifiers! Each of the classifiers can be developed in a different context and for entirely different representation/description (identification of persons by their voice, face as well as hand writing) More than a single training set is available

65 Combine classifiers! Different classifiers trained on the same data may only differ in their global performances May show strong local differences Combination is very useful if the individual classifiers are independent

66 Combination Parallel All classifiers are evaluated independently Their results are combined (first weight then combine!!!) Serial - linear sequence The number of possible classes for a given pattern are gradually reduced as more classifiers in the sequence has be involved First consider simple classifiers (low computational and measurements cost) More accurate and expensive classifiers apply after Hierarchical (tree-like) Individual classifiers are combined into a structure (a tree-like structure) Advantage high efficiently and flexibility

67 Unsupervised classification Difficult, expensive or even impossible to rely label training sample with its true category (It is not possible to obtain ground truth) Patterns within a cluster are more similar to each other than are patterns belonging to different clusters

68 Unsupervised classification Difficult How to determine the number of clusters K? The number of clusters often depends on resolution (fine vs. coarse)

69 K means Step 1. Select an initial partition with K clusters. Repeat steps 2 through 4 until the cluster membership stabilizes Step 2. Generate a new partition by assigning each pattern to its closest cluster center Step 3. Compute new cluster centres as the centroids of the clusters Step 4. Repeat steps 2 and 3 until an optimum value of the criterion function is found (the cluster are stabilized)

70 K means

71 K means

72 K means - disadvantages Different initialization can result in different final clusters Fixed number of clusters can make it difficult to predict what K should be It is helpful to rerun the classification using the same as well as different K values, to compare the achieved results 10 different initializations for 2D data For N-dimensional data 10 different initializations is often not enough!

73 How to determine K? Davies-Bouldin index (DB index) It works for spherical clusters Intra-cluster and inter-cluster measure

74 Hierarchical clustering Hierarchical clustering - construct clustering tree (dendrogram) Start with each object/pixel as its own class Merge the classes that are closest according to some distance measure Continue until only one class is achieved Decide the number of classes based on the distances in the tree

75 Simple dendrogram

76 Distances measure and linking rules Distance measures: Euclidean City block Mahalonobis Linking rules - how distances are measured when cluster contains more than one object Single linkage (nearest neighbours) Complete linkage (furthest neighbours) Mean linkage (distance between cluster centres)

77 Linking rules Dependent on metric d complete - red d single - blue

78 Summary and conclusions Classification is needed in image processing It is highly application dependent Features and classifiers Using more features does not guarantee a better result!

Classification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University

Classification. Vladimir Curic. Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University Classification Vladimir Curic Centre for Image Analysis Swedish University of Agricultural Sciences Uppsala University Outline An overview on classification Basics of classification How to choose appropriate