Visual Processing in Fingerprint Experts and Novices Tom Busey Indiana University, Bloomington John Vanderkolk Indiana State Police, Fort Wayne Expertise with fingerprint examiners was tested in behavioral and EEG studies. Experts show greater tolerance for noise, are unaffected by longer memory delays, and show evidence of configural processing. This last finding was confirmed in an EEG study where experts show a reliable delay of the N170 component when fingerprints were inverted, while novices did not. Configural processing may be one element that underlies perceptual expertise. www.indiana.edu/~busey/
How Do Experts Make Identifications? Easy Match Hard Match
What Cognitive Abilities Support Expertise? Learn relevant features or dimensions Recognize new clusters of features (unitization) Robustness to noise Tune detectors to specific characteristics of features (exclude noise) Integrate information over larger regions of space Learn to trade off quantity and quality in decision making Superior visual memory Tolerate affine transformations such as rotation and stretch
Study Fragment one second
Mask Either 200 ms or 5200 ms
Test Images Until Response
Testing Fingerprint Expertise: X-AB Sequential Matching Task example stimulus pairs:
Reduce Matching based on Low-Level Features Overall Brightness change Study image is rotated up to 90° in either direction Two image manipulations designed to simulate latent prints Added noise Partial masking
Added Noise
Partial Masking
Partial Masking Semi-Transparent Masks Fingerprint Partially Masked original inverse Logical Combination Recovers Original Fingerprint Semi-Transparent Masks Fingerprint Partially Masked Fingerprints
Includes combinations:
Image Degradations at Test
Partial Masking Semi-Transparent Masks Fingerprint Partially Masked original inverse Summation Recovers Original Fingerprint Semi-Transparent Masks Fingerprint Partially Masked Fingerprints
Partial Images in Noise Behavioral Data Full Images Partial Images Full Images in Noise Partial Images in Noise Experts: No effect of delay, interaction between noise and partial masking.
Logic of Partial Masking One Half Both Halves One Half Logical Combination Recovers Original Fingerprint Partially Masked Fingerprints
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) Correct Decision d = likelihood of getting enough information from half of a print in order to make a correct decision db: d when both halves are present
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? yes (db) no (1-db) Correct Decision d = likelihood of getting enough information from half of a print in order to make a correct decision db: d when both halves are present
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? yes (db) no (1-db) info from guessing? yes (g) no (1-g) Correct Decision Wrong Decision d = likelihood of getting enough information from half of a print in order to make a correct decision db: d when both halves are present g = getting the correct answer by guessing (high threshold model)
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? info from first half? yes (db) no (1-db) yes (do) no (1-do) info from guessing? info from guessing? yes (g) no (1-g) yes (g) no (1-g) Correct Decision Wrong Decision Correct Decision Wrong Decision d = likelihood of getting enough information from half of a print in order to make a correct decision db: d when both halves are present do: d when only one half present g = getting the correct answer by guessing (high threshold model)
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? info from first half? yes (db) no (1-db) yes (do) no (1-do) info from guessing? info from guessing? yes (g) no (1-g) yes (g) no (1-g) Correct Decision Wrong Decision Correct Decision Wrong Decision Question: What is the relation between db and do? db = do : One half doesn't influence information acquired from other half db < do : Get less information from one half when second is present db > do : Get more information from one half when second is present (consistent with configural or gestalt processing)
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? info from first half? yes (db) no (1-db) yes (do) no (1-do) info from guessing? info from guessing? yes (g) no (1-g) yes (g) no (1-g) Correct Decision Wrong Decision Correct Decision Wrong Decision Testing: Fit reduced model with db = do. Can we reject this model? If we reject it, look at the relation of db to do by fitting the full model. db >> do is consistent with configural processing.
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? info from first half? yes (db) no (1-db) yes (do) no (1-do) info from guessing? info from guessing? yes (g) no (1-g) yes (g) no (1-g) Correct Decision Wrong Decision Correct Decision Wrong Decision Experts: No noise: reject reduced model, db and do are significantly different Full model: db = .841, do = .944 wrong direction for configural processing In noise: reject reduced model, db and do are significantly different Full model: db = .50, do = .30 consistent with configural processing
Multinomial Tree Modeling Full Image (Both Halves) Partial Image (One Half) info from first half? yes (db) no (1-db) info from second half? info from first half? yes (db) no (1-db) yes (do) no (1-do) info from guessing? info from guessing? yes (g) no (1-g) yes (g) no (1-g) Correct Decision Wrong Decision Correct Decision Wrong Decision Novices: No noise: reject reduced model, db and do are significantly different Full model: db = .40, do = .54 wrong direction for configural processing In noise: can't reject reduced model, db and do are not significantly different Full model: db = .19, do = .13 No evidence for configural processing
Evidence for Configural Processing: Multinomial Modeling To test for configural processing, we can use the accuracy rate in the partial image condition to make a prediction for the full image condition, assuming no configural processing. If performance in the full image condition exceeds the prediction, we have evidence that is consistent with configural processing.
Evidence for Configural Processing: Multinomial Modeling To test for configural processing, we can use the accuracy rate in the partial image condition to make a prediction for the full image condition, assuming no configural processing. If performance in the full image condition exceeds the prediction, we have evidence that is consistent with configural processing. Experts in noise: We predict performance in the full image condition to be about 75% correct. Instead it is around 90%. Experts are doing better with the whole image than we predict they would do based on partial-image performance. This is configural processing at work.
Configural Processing in Faces: The ‘Thatcher Illusion’ Features are perceived individually, image looks ok. (Thomson, 1980)
Configural Processing in Faces: The ‘Thatcher Illusion’ Features are perceived individually, image looks ok. Features are perceived in context, image looks grotesque. (Thomson, 1980)
Configural Processing in Faces: The ‘Thatcher Illusion’ Features are perceived individually, image looks ok. Features are perceived in context, image looks grotesque. (Thomson, 1980)
Configural Processing in Faces: The ‘Thatcher Illusion’ Features are perceived individually, image looks ok. Features are perceived in context, image looks grotesque. (Thomson, 1980)
EEG Recording Basics Record from the surface of the scalp Amplify 20,000 times Electrical signals are related to neuronal firing, mainly in post-synaptic potentials in cortex. Very small signals, very noisy data.
EEG Recording Basics Average over lots of trials (200 trials per condition)
EEG and Configural Processing Faces produce a strong component over the right hemisphere at about 170 ms after stimulus onset, which is called the N170. Inverted faces cause a delay of 10-20 ms in the N170. Trained objects (Greebles) show a delay in the N170 component with inversion, but only in the left hemisphere (channel T5). Data from Rossion, Gauthier, Tarr, Despland, Bruyer, Linotte & Crommelinck (2000) fMRI studies show IT is active whether attending to faces or not (Tarr) No effect of familiarity (Bentin & Deouell, 2000) target status: Is the face-sensitive N170 the only ERP not affected by selective attention? (Caquil, Edmonds, & Taylor, 2000) Appears to be feed-forward perceptual processing of faces or other face-like stimuli Coupled with behavioral data suggesting configural processing with faces, an advanced N170 to an upright stimulus suggests that the N170 latency differences indicate configural processing. Data from Rossion, Gauthier, Goffaux, Tarr & Crommelinck (2002)
An Obvious Experiment: Show upright and inverted fingerprints to Fingerprint examiners and novices. If experts process fingerprints configurally, we should see a delayed N170 to inverted fingerprints. fMRI studies show IT is active whether attending to faces or not (Tarr) No effect of familiarity (Bentin & Deouell, 2000) target status: Is the face-sensitive N170 the only ERP not affected by selective attention? (Caquil, Edmonds, & Taylor, 2000) Appears to be feed-forward perceptual processing of faces or other face-like stimuli Also test faces to replicate the face inversion effect in our subjects. Test both identification and categorization tasks.
Summary of Experiment Fingerprint experts demonstrate strong performance in an X-AB matching task, robustness to noise and evidence for configural processing when stimuli are presented in noise. This latter finding was confirmed using upright and inverted fingerprints in an EEG experiment. Experts showed a delayed N170 component for inverted fingerprints in the same channel that they show a delayed N170 for inverted faces. Thus they appear to be processing upright fingerprints in part using configural or holistic processing, which stresses relational information and implies dependencies between individual features. In the case of fingerprints, this may come from idiosyncratic feature elements instead of well-defined features such as eyes and mouths. www.indiana.edu/~busey/