Transcript PatReco: Introduction

PatReco: Detection
Alexandros Potamianos
Dept of ECE, Tech. Univ. of Crete
Fall 2009-2010
Detection
 Classification problems with two classes are
sometimes referred to as detection problems
 For detection problems the two classes are
referred to as ω2 and ω1 = NOT ω2
 In statistics NOT ω2 is the null hypothesis H0
and ω2 is H1
Detection
 Goal: Detect an Event




Hit (Success): event occurs and is detected
False Alarm: event does not occur but is detected
Miss (Failure): event occurs and goes undetected
Correct Reject: event neither occurs nor detected
 In traditional Bayes classifier terms:


P(correct) = P(Hit) + P(Correct Reject)
P(error) = P(False Alarm) + P (Miss)
Detection Examples
 House Alarm (detect burglary)
 Reading bits of a CD or a DVD (detect 1’s)
 Medical screening (e.g., detect cancer)




Hit (Success): cancer present and detected
False Alarm: caner not present but not detected
Miss (Failure): cancer present and goes undetected
Correct Reject: cancer neither present nor detected
Further
testing
No
action
Miss
Correct
Reject
Hit
False
Alarm
Receiver Operator Curve (ROC-curve)
Equal Error Rate(EER)
Operation Point
Conclusions
 Detection is a special case of two-class
classification
 Type I and Type II errors (miss and false alarms)
often have different costs
 Often increase Bayes error to minimize total cost
 Select an operation point on the ROC-curve