de-identification of facial images by use of composites *mark e. engelstad md, dds, mhi oregon...
TRANSCRIPT
De-identification of Facial Images by Use of Composites
*Mark E. Engelstad MD, DDS, MHIOregon Health & Science UniversityDept of Oral and Maxillofacial SurgeryDept Medical Informatics & Clinical Epidemiology
Genevieve B. Melton, MD, MAUniversity of MinnesotaInstitute for Health InformaticsDepartment of Surgery
Medbiquitous Annual Symposium, Baltimore MD May 10, 2011
The Questions:
Do composites de-identify faces?
Even those that are well-known to an observer?
Are facial composites realistic in appearance?
Figure 2: A comparison of two techniques for facial image de-identification. The middle image (B) is the original image. (A) black boxes only. (C) a facial composite, altered in the area of eyes and eyebrows only.
Me
Faces
Subjects viewed the composite faces twice—first unaware that the faces were composites, and then primed to the presence of composites.
• Subjects viewed 20 composite faces
• 10/20 had a third of a familiar face (test face)
Test Face
Test Face
Results
Subject Response Unprimed (1st Viewing)
Primed (2nd Viewing)
Facial CompositesTotal = 20
Composites of Unfamiliar
Faces Total = 10
Did Not Identify (True Neg)
100% (120/120 ) *
42% (50/120)
Identified Wrongly(False Pos)
0% (0/120) 58% (70/120)
Composites with Familiar (Test) Faces
Total = 10
Identified Correctly (True Pos)
0% (0/120) 62% (74/120)
Identified Wrongly (False Pos)
0% (0/120) 19% (23/120)
Failed to Identify (False Neg)
100% (120/120 ) *
19% (23/120)
No subjects identified test faces unless they were primed to their presence (* p < 0.001).
Results
Familiar Face Composite A Region Visible
Familiar Face Composite B Region Visible
Faces A and B Views by region (n)
42%(5/12)
79% *(19/24)
67%(24/36)
Upper
36
71% †(17/24)
38% †9/24
54%(26/48)
MidFace
48
67%(8/12)
67%16/24
67%(24/36)
Lower
36
Total Face A
63%(30/48)
Total Face B
61%(44/72)
Total Faces A and B62%
(74/120)
Total
120
Table 2: Identification of Test faces after priming--compared by facial region. Percentages of subjects who correctly identified a familiar face when regions of that face were visible in the composite image are shown (true positives). In Test Face B, a significant difference (* p<0.01) in identification rate existed between Upper Face and Midface. Test Face A Midface was recognized correctly more often than Test Face B Midface (†p<0.01)
4: Size all images to a standard (800 x1200)
5: Align the facial features
6: Create a Layer Mask
7: Use a Brush to reveal deeper layer
8: Blend the edges between the two layers
9: Correct Color Tones