人工智能與模式識(shí)別的發(fā)展

1、1,1,Some Recent Development of Intelligent PR and Applications,Guanghui He ghhe@cqu.edu.cn,2,2,What are Biometrics?Biometrics are automated methods of recognizing a person based on the acquired physiological or behavi

2、oral characteristics,Percentage of usage (Source: International biometric group),3,3,A ScenarioTwo Al Qaeda（“基地”組織） suspects were recently taken into custody by U.S. immigration authorities as they tried to enter the Un

3、ited States after their fingerprints were matched with ones lifted by U.S. military officials from documents found in caves in Afghanistan（阿富汗）.,Why Biometric Technologies?For Security Reasons,4,4,Example 1: SFinGe - S

4、ynthetic Fingerprint Generator developed at the Biometric Systems Lab,University of Bologna – ITALY, is utilized to:,compare different fingerprint matching algorithms,train pattern recognition techniques that require l

5、arge learning-sets (e.g. neural network),easily generate a large number of “virtual users” to develop and test medium/large-scale fingerprint-based systems,5,5,,,3-D model (pressure in on-line model),Modeling by deformat

6、ion,Modeling segments (conics, splines),Example 2: generation of synthetic signature,Assembling (desegmentation) of 2-D model,6,6,Example 3: Privacy protection: After enrollment, a true object (e.g. image of fa

7、ce, fingerprint or voice signal) is intentionally distorted using irreversible transform - Cancelable biometrics (Ratha, Connell, Bolle, 2001),Skin distortion (fingerprint) (source: Biometric Systems Lab, University o

8、f Bologna),Face image is warped with bilinear interpolation (source: Serif Inc.),Some More Examples: Generation of synthesis fingerprints Generation of synthetic signatures (handwriting modeling is a relevant problem)

9、 Iris recognition and synthesis Information fusion in biometrics Speech-to-animated-face,7,8,Where do we need biometrics?,,Traditional application: human identification Recent advances: Early warning paradigm D

10、esigning simulators for HQP training systems Sensing in robotics,9,Early detection and warning,,,,,,Semantic domain,,Biometric sensor,Signal processing,Decision making,,,,,Raw biometric data,,,Basic configuration,F

11、eature space,,,,Application: physical access control system,Sensors,ExtractorsImage- andsignal processingalgorithm,Classifiers,,,BiometricsVoice, signature, face, fingerprint, iris, hand geometry, etc,Data Rep.A

12、udio signal, image, infrared image,FeatureVectors,Scores,,,,,,,Decision:Match, Non-match,Inconclusive,,Biometric databases,Level 1: document-check,Databases (Watch-list),,,Level 2: biometrics,10,11,Laboratory exper

13、iments,,,12,Early warning system components:,- Supports facial analysis Skin temperature evaluation Detection of disguise: wig and other artificial materials, and surgical alternations Evaluation of blood vessel flow

14、 (modeling expressions) Other physiological / medical measurements (alcohol / drug abuse),Infrared biometrics and decision support,Mid-infrared: 3-5 ??m, far-infrared: 8-12??m,Temperature value 32.8754 0C i

15、s detected in a point,13,Early warning system components:,Blood flow rate analysis (from infrared),Visualization of the blood flow rate from the upper rectangle of (a),Thermal image of subject at the beginning of answer

16、ing the question “Do you have that stolen $20 on you right now?”,Thermal image of subject at the end of answering the question,Visualization of the blood flow rate from (b). The difference is significant (from I. Pavl

17、idis’ report),14,Early warning system: decision-making,,insufficiency of information,INDIVIDUALbiometrics,Degrees of belief,Biometric sensor,,TEMPORAL faults of biometric sensors,,errors of biometric sensors,Mass assig

18、nments,,,,,,,,,Belief function,,,,,,,Updating,,Decision making in semantic form,15,Early warning security access control system:,Semantic processor,Gait-biometric processor,,Gait features processor,,,The ground reaction

19、 force,GenderPregnancyFatigueInjuriesAfflictionsDrunkenness,,,,,,,,,Ground reaction force processor,,,,,,,Discriminative gait biometric in semantic form,Gait biometrics analysis and decision-making assistance,16,Fac

20、e capturing,Fitting points,,,,0001001001010011010010010010010110010010001000010010110100100101001001001000 …,File (mesh/colour),3D Face model,Early warning system components:,17,Face capturing,Fitting poi

21、nts,,,,0001001001010011010010010010010110010010001000010010110100100101001001001000 …,File (mesh/colour),3D Face model,Early warning system components:,18,Other applications:Biometric data modeling for H

22、QP training,,Processing of screened data,,Processing of pre-screeneddata,,,,,,,,Dialogsupport,Decision-making support,,,,Visible band camera,IR band camera,Synthetic image of an individual,Voice analyzer,,Officer-in-tr

23、aining,19,Perspectives: humanoid robots,,Sensing in robotics,Robot head developed by Dr. Marek Perkowskiat Portland State University,Emotion synthesis,Robot speech,20,20,It’s Similarity and Pattern Matching!,What is Mea

24、surement ?,Just a Comics Joke?,No! More Than That,21,21,Pattern Recognition,Cognition (Learning)Re-CognitionClassificationIdentificationVerificationClustering,22,22,3D Object Recognition,23,23,Table of Contents,BACK

25、GROUNDTHEORYEXPERIMENTS and ILLUSTRATIONSFUTURE RESEARCH,24,24,Linear Combination,Object 1 A1Object 2 A2Object 3 A3Object 4 A4Object A4= a A1+ bA2 +cA3 +d,25,25,3D Recognition Background,Widely usedindustrial par

26、ts inspection military target identificationCAM/CAD engineering design image/vision understanding, interpretation, visualization, and recognition,26,26,3D Recognition Background,Recognition 3D objectsRigid Objects

27、 Fixed shapesDeformable Objects Variable shapesArticulated Objects Fewer methods proposed Brooks' ACRONYM system using symbolic reasoning. Grimson et al extended the interpretati

28、on of tree approach to deal with 2-D objects with articulated components,27,27,3D Recognition Background,Extended Linear Combination Method (LC)Simpler preprocessing Simpler and faster computation Applicable to man

29、y articulated object recognition, understanding, interpretation, and visualization,28,28,THEORY,Extended Linear Combination Method (LC)based on the observation that novel views of objects can be expressed as linear comb

30、ination of the stored views (from learning) It identifies objects by constructing custom-tailored templates from stored two-dimensional image models.,29,29,Linear Combination,Modelan image consists of a list of feature

31、 points observed in the image,30,30,Linear Combination,,Recognition: An unknown object is matched with a model by comparing the points in an image of the unknown object with a template-like collection of points pr

32、oduced from the model,31,31,32,32,33,33,34,34,35,35,Experitment-1——Match same objects,,,36,36,Experiment-1 Result,,37,37,Experiment-2,,,38,38,Experiment-3 …,,39,39,Experiment-3 Result,,,40,40,Experiment-4,,41,41,Experime

33、nt-4 Result,,,Rejected,Rejected Too,,,42,42,43,43,44,44,45,45,46,46,Color Biometric Imaging Analysis,47,47,Items to be discussed:,Clustering and K-means algorithmStatisticalUnsupervisedColor Representation and Color I

34、mage Segmentation,48,48,Supervised Classification and minimum distance classification,Minimum Distance ClassificationSupervisedFind the center of known patterns of each class Classify unknown patterns into the clas

35、s that is “closest” to it.,,,49,49,Color Image Segmentation: Hue Component,50,50,Color Image Segmentation,Task:Study the K-means algorithm in hue space.Interesting: Periodical Circular Property of hue componentnew M

36、easure of Distance.Problem:K-means algorithm is based on the measure of distance and definition of center,51,51,Hue Component Clustering,Definition 1: Distance of Hue ValuesDefinition 2: Directed Distance of Hue Value

37、sTricky: Addition of Directed DistanceDefinition 3: Interval and Its Midpoint in H Space.Definition 4: Center of a Set of Points in Hue SpaceTheory: Euclidean Theory of Center in Hue Space,52,52,Hue Component Cluster

38、ing,Definition 1: Distance of Hue Values,,53,53,Hue Component Clustering,Definition 2: Directed Distance of Hue ValuesTricky: Addition of Directed Distancethe following vector addition property no longer holds:,,,,5

39、4,54,Hue Component Clustering,Revisit definition: Interval and Its Midpoint in H Space.Revisit definition : Center of a Set of Points in Hue SpaceRevisit the Proof of Theory: Euclidean Theory of Center in Hue Space,55,

40、55,Color Image Segmentation,I and H components are of Interest. Good color image segmentation algorithms should consider and combine bothVariation of light intensity and occlusion: hue component is betterColor info

41、rmation is lost:Intensity component is better Fuzzy member function is introduced,56,56,Color Image Segmentation - Experiment 1Intensity Distinguishable,(a) Original color image,57,57,Color Image Segmentation - Exper

42、iment 1Intensity Distinguishable,(b) Intensity image,58,58,Color Image Segmentation - Experiment 1Intensity Distinguishable,(c) Hue image,59,59,Color Image Segmentation – Experiment Intensity Distinguishable,(d) Segmen

43、tation by hue,60,60,Color Image Segmentation - Experiment 1Intensity Distinguishable,(e) Segmentation by hue and intensity,61,61,Color Image Segmentation - Experiment 2 Hue Distinguishable,(a) Original color image,62,6

44、2,Color Image Segmentation - Experiment 2 Hue Distinguishable,(b) Intensity image,63,63,Color Image Segmentation - Experiment 2 Hue Distinguishable,(c) Hue image,64,64,Color Image Segmentation - Experiment 2 Hue Disti

45、nguishable,(d) Segmentation by intensity,65,65,Color Image Segmentation - Experiment 2 Hue Distinguishable,(e) Segmentation by hue and intensity,66,66,Some More Illustrative Examplesof Medical Imaging Results,67,67,68,