Controlling misses and false alarms in a machine

Controlling misses and false alarms in a machine learning framework for predicting uniformity of printed pages February 2015 Ph. D. student: MINH Q NGUYEN (nguyen 40@purdue. edu) Professor: JAN P ALLEBACH (allebach@purdue. edu) Purdue University Minh Q Nguyen & Jan P Allebach – EI 2015 1

Outline 1. Review our previous work 2. Motivation 3. Theory of Support Vector Machine (SVM) 4. Result & Discussion Minh Q Nguyen & Jan P Allebach – EI 2015 2

0. 25 inch Grain Mottle Nondirectional Nonuniformity Print Defects Wood grain/Tiger Marks Minh Q Nguyen & Jan P Allebach – EI 2015 Scanned at 600 dpi Pin holes 3

Algorithm l 6 models to characterize uniform print quality * *: Minh Q. Nguyen, Renee Jessome, Stephen Astling, Eric Maggard, Terry Nelson, Mark Shaw, Jan P. Allebach, "Perceptual metrics and visualization tools for evaluation of page uniformity", Proceedings of SPIE Vol. 9016, 901608 (2014) MDE-Intra. BF: MDL-Intra. BF : DDE-Intra. BF : DDL-Intra. BF : SDE-Inter. BF : L*/a*/b*-Inter. BF: Minh Q Nguyen & Jan P Allebach – EI 2015 Mean ∆E Intra-Block Fluctuation Mean ∆L Intra-Block Fluctuation Dispersion of ∆E Intra-Block Fluctuation Dispersion of ∆L Intra-Block Fluctuation Signed ∆E Inter-Block Fluctuation L*/a*/b* Inter-Block Fluctuation 4

Graininess MDE MDL DDE DDL SDE L* a* b* Nonuniformity Minh Q Nguyen & Jan P Allebach – EI 2015 5

Support Vector Machine Red pages Minh Q Nguyen & Jan P Allebach – EI 2015 Cyan pages 6

Motivation l SVM typically solves for maximal accuracy. In many applications, the weights of misses and false alarms are not always equal l A "miss" : a page of unacceptable quality but the algorithm declares to be a "pass". “False alarm”: a page of acceptable quality, but is flagged by the prediction algorithm as a "fail” l "False alarm" pages result in extra pages to be examined, which increases labor cost. Misses are a serious problem, since they represent problems that will not be seen by the systems designers l This scenario motivates us to develop a machine learning framework that will achieve the minimum "false alarm" rate subject to a specified "miss" rate Minh Q Nguyen & Jan P Allebach – EI 2015 7

SVM for Linearly Separable Classes Linear model: Minh Q Nguyen & Jan P Allebach – EI 2015 8

SVM for Non-linearly Separable Classes Minh Q Nguyen & Jan P Allebach – EI 2015 9

Controlling Misses and False Alarms with SVM * * Minh Q Nguyen & Jan P Allebach – EI 2015 10

Parameter C l C controls the tradeoff between slack variable penalty and margin. l Large C: narrow separation zone with fewer support vectors. Small C: the separation zone becomes larger and has more support vectors involving in the decision making C = 10 C=1 Minh Q Nguyen & Jan P Allebach – EI 2015 11

Minh Q Nguyen & Jan P Allebach – EI 2015 12

No misses No False Alarms Minh Q Nguyen & Jan P Allebach – EI 2015 13

Tuning all parameters • Perform a grid search all over the space of • Update the table if the new FA rate is less than the FA rate in the table with the same Miss rate. 0% 1% LUT 2% 3% … … … 99% Minh Q Nguyen 100% & Jan P Allebach – EI 2015 14

Constructing ROC curves Cyan pages Minh Q Nguyen & Jan P Allebach – EI 2015 Magenta pages 15 Red pages

Summary l Introduce a SVM framework to control misses and false alarms for predicting print uniformity l Several parameters need to be tuned to formulate a LUT in which each miss rate corresponds to a minimum false alarm with specific values of l Construct ROC curves for practical application Minh Q Nguyen & Jan P Allebach – EI 2015 16

Thank you for your attention! nguyen 40@purdue. edu allebach@purdue. edu Minh Q Nguyen & Jan P Allebach – EI 2015 17