Pixel Counting Biometric Comparison of Oswald photos

[Drew Phipps]

This will be rather long, and probably broken up into a few posts.

Digitally Comparing Oswald photos
Drew Phipps 2015

I am going to use my pixel counting software (AnalyzingDigitalImages) to measure ratios of distance across and around the face of the various pictures of Oswald as a boy and as a man. The Warren Commission, and the HSCA, said that all these pictures are of the same person. John Armstrong, in his comprehensively researched tome "Harvey and Lee" says there are two separate individuals. We will see if pixel counting can resolve this issue.

As a control measure, I am also going to pixel count some other familiar faces from the same time period. That should give us some idea of what is a significant difference in these biometric ratios. The ratios I will use are: "pupil-to-pupil / width of eye" (called P/W hereafter), "pupil-to-pupil / length of nose" (called P/N hereafter), "pupil-to-pupil / nose-to-top-lip" (called P/L hereafter) and "pupil-to-pupil / earlobe-to-earlobe" (called P/E hereafter). The use of ratios (instead of actual measurements) will make it unnecessary to know more about the distance from lens to face, or the type of camera, etc., since the proportions of the face of the same person should stay the same regardless of those other factors.

I don't claim that these are the best biometric ratios to use, I frankly don't know which are the best. These seem to me the simplest to acquire. I also guess that measurements to the lower part of the face involving the jaw might vary greatly depending on facial expression, so I'm not going to use that part of the face. It is also said that the nose and the ears keep growing as people age.

(Warning about pixel counting: As in every other form of digitization, when a picture is digitized, information is "smeared" across each pixel. Therefore, you can never be sure if the actual boundary of an object is just a little bit into the pixel, halfway across the pixel, or most of the way across the pixel. That adds a certain degree of uncertainty. The more pixels in an image, the "better" the certainty of the measurement, and fewer pixels gives greater uncertainty. The photos I'm comparing were taken in the 1960's and the digital versions are low definition.)

[Drew Phipps]

Here is a drawing illustrating the problem of pixel measurements. Since I am making linear ratio comparisons (x1/x2), it's more akin to an area problem than a linear measurement problem:


The yellow circle has an area of less than 1 square; however, it would probably show up after pixilation as 1 square unit. The red circle has an actual area of more than 1 square unit (1.57); after pixilation it could show up as anything from 1 square unit to 5 square units. The blue circle has an actual area of 4.9, after pixelation, it might be anything from 1 to 9 units (but most likely around 5). The more pixels that are present in a picture, the less significant the error becomes.

Recent government studies on pixelation (in ultrasound images for medical purposes) seem to suggest the proper error factor is 1.5 pixels for a linear measurement. Since what I am comparing here is ratios, an expected error factor is (1.5)^2 = 2.25 sq. pixels. Any variations of ratio between photos that is less than 2.25 sq. pixels, I'm going to call insignificant, and any variation greater than 2.25 sq. pixels, I'm going to call significant. ( There's probably a more specific and scientific way to determining the significance of differing ratios, but I don't know it. )

In addition, in comparing ratios of whole numbers whose values are less than 100 (in most cases), the data may be "too lumpy," (to borrow a phrase) and therefore produce apparently significant results where none really exist. (I.E. the difference between 3/5 and (3+1)/5 is far greater than the difference between 300/500 and (300+1)/500). Unfortunately the low-resolution aspect of these photos forces me to work with low numbers.

[Drew Phipps]

CONTROL COMPARISONS:

Here is our control group:

First, the pictures:

1. Booking pic of Oswald
(this will be the baseline for comparing all the Oswald photos)



2. Booking pic of Ruby


3. Pic of JFK


4. Pic of Jackie


5 Pic of Ferrie




(Measurements are (in order) eye width, pupil to pupil, nose length, nose to lip, earlobe to earlobe)

Photo Measurements

Booking pic of Oswald : 20, 46, 37, 15, 97 Error = 4.9% +0

P/W .435 P/N 1.243 P/L 3.067 P/E .474

Booking pic of Ruby: 22, 54, 45, 18, 127 Error = 4.2%+4.9%

P/W .407 P/N 1.200 P/L 3.000 P/E .425

Pic of JFK: 164, 357, 269, 134, 808 Error = 0.6%+4.9%

P/W .459 P/N 1.327 P/L 2.664 P/E .442

Pic of Jackie: 91, 253, 173, 51, 469 Error = 0.9%+4.9%

P/W .360 P/N 1.462 P/L 4.961 P/E .539

Pic of Ferrie : 35, 89, 73, 24, 190 Error = 2.5%+4.9%

P/W .393 P/N 1.219 P/L 3.708 P/E .468

As you can see, the higher resolution photos contain far less uncertainty.


Comparison of control group as a percent of Oswald's ratios
(significant difference underlined)

Oswald +-4.9%

P/W 100% P/N 100% P/L 100% P/E 100%

Ruby +-9.1%

P/W 93.6% P/N 96.5% P/L 97.8% P/E 89.7%

JFK +-5.5%

P/W 105.5% P/N 110.2% P/L 86.9% P/E 93.2%

Jackie +-5.8%

P/W 82.8% P/N 117.6% P/L 161.8% P/E 113.7%


Ferrie +-7.4%

P/W 90.3% P/N 98.1% P/L 120.9% P/E 98.7%



As you can see, the individual variations from Oswald's biometrics, compared to different individuals, were no less than 1.3%, and no greater than 61.8% However, each individual had at least one ratio that was significantly different than Oswald, even when the pixel error factor was considered. (Ruby, strangely, resembles Oswald, at least mathematically).

[Drew Phipps]

Now for our "experimental" group of test photos:

1. Booking pic of Oswald

2. Booking NOLA



3. Oswald USMC


4. Oswald Marine uniform

5. Oswald Marines



6. Oswald Minsk


7. Oswald crew cut


8. Oswald teenager



9. Oswald at zoo


10. Oswald schoolboy


(more on next panel)

[Drew Phipps]

11. Oswald NOLA cadet

12. Oswald from Lee & Me


13. Multiple kid (avg. of 3 full frontal shots)



Error factor is 2.25 sq. pixels/ (pupil-to-pupil measurement) = xyz % , since I use the pupil-to-pupil measurement in each of the proportions.

MEASUREMENTS:

Booking pic of Oswald: 20, 46, 37, 15, 97 Error=4.9%

P/W .407 P/N 1.243 P/L 3.067 P/E .474

Booking NOLA: 16, 37, 29, 14, 78 Error=4.9%+6.1%=11%

P/W .432 P/N 1.276 P/L 2.642 P/E .474

Oswald USMC: 20, 51, 41, 14, 108 Error=4.9%+4.4%=9.3%

P/W .392 P/N 1.244 P/L 3.643 P/E .472

Oswald Marine uniform: 17, 38, 31, 14, 76 Error=4.9%+5.9%=10.8%

P/W .447 P/N 1.225 P/L 2.214 P/E .500

Oswald Marines 17, 43, 34, 18, 89 Error=4.9%+5.2%=10.1%

P/W .395 P/N 1.264 P/L 2.389 P/E .483

Oswald Minsk: 16, 42, 30, 17, 91 Error=4.9%+5.4%=10.3%

P/W .381 P/N 1.400 P/L 2.471 P/E .462

Oswald crew cut: 16, 36, 28, 14, 77 Error=4.9%+6.3%=11.2%

P/W .444 P/N 1.856 P/L 2.571 P/E .468

Oswald teenager: 17, 48, 39, 18, 101 Error=4.9%+4.7%=9.6%

P/W .354 P/N 1.231 P/L 2.667 P/E .475

Oswald at zoo: 10, 22, 16, 8, 50 Error=4.9%+10.2%=15.1%

P/W .454 P/N 1.375 P/L 2.750 P/E .440

Oswald schoolboy 19, 40, 29, 16, 85 Error=4.9%+5.6%=10.5%

P/W .475 P/N 1.379 P/L 2.500 P/E .471

Oswald NOLA cadet: 13, 30, 22, 12, 62 Error=4.9%+7.5%=12.4%

P/W .433 P/N 1.363 P/L 2.500 P/E .471

Oswald from Lee and Me 14, 34, 25, 13, 71 Error=4.9%+6.6%=11.5%

P/W .411 P/N 1.360 P/L 2.615 P/E .479

Multiple images (avg 3 shots) 21, 55, 41, 17, 117 Error=4.9%+4.1%=9.0%

P/W .381 P/N 1.341 P/L 3.236 P/E .470



Percent Comparisons between Oswald booking photo: '
(significant difference underlined)

Oswald

P/W 100% P/N 100% P/L 100% P/E 100%
+-4.9%

Booking NOLA:

P/W 106.1% P/N 97.4% P/L 86.1% P/E 100%
+-11.0%

Oswald USMC

P/W 96.3% P/N 99.9% P/L 84.2% P/E 100%
+-9.3%
Oswald Marine uniform:

P/W 109.8% P/N 101.5% P/L 138.5% P/E 98.1%
+-10.8%

Oswald Marines:

P/W 97.1% P/N 98.3% P/L 128.4% P/E 98.1%
+-10.1%

Oswald Minsk:

P/W 93.6% P/N 88.8% P/L 124.1% P/E 102.6%
+-10.3%

Oswald crew cut:

P/W 109.0% P/N 96.7% P/L 119.3% P/E 101.3%
+-11.2%

Oswald teenager:

P/W 87.0% P/N 101.0% P/L 115.0% P/E 100.0%
+-9.6%

Oswald at zoo:

P/W 111.5% P/N 90.4% P/L 111.5% P/E 107.7%
+-15.1%


Oswald schoolboy:

P/W 116.7% P/N 90.1 P/L 122.7% P/E 100.6%
+-10.5%

Oswald NOLA cadet:

P/W 106.4% P/N 91.2% P/L 122.7% P/E 100.6%
+-12.4%

Oswald from Lee and Me

P/W 100.1% P/N 91.4% P/L 117.3% P/E 99.0%
+-11.5%

Multiple images (average 3 frontal shots)

P/W 93.6% P/N 92.7% P/L 94.8% P/E 100.9%
+-9.0%

[Drew Phipps]

Preliminary Conclusions:

The pupil/lip ratio, especially, appears to be way out of whack between the Dallas Booking photo of Oswald (11/22/63) and every other photo of Oswald (with the exception of the zoo and the kiddie shots in the multiple image picture).


This also could indicate a different person, could be a failure of the methodology I have used in this comparison, or it could be something wrong with Oswald's upper lip after his arrest (it seems to be too short). I can't figure out what facial expression, besides smiling (which he isn't doing in the booking photo), could shorten the distance in the booking photo from nose to upper lip. There are several which might lengthen the distance, but shortening that distance seems problematic. Was Oswald's lip swollen from his arrest? Looking at profile shot after his arrest, it doesn't appear so.

The only other apparently significant differences is that the "Minsk" Oswald's nose is too long, the "Teenage" Oswald's pupils aren't wide enough (which might actually be a result of the squinting), and the "Schoolboy" Oswald's pupil's are too wide.

By contrast, the pupil/earlobe ratio appears to fall within the range of error due to pixelation for every one of these photos. That seems to indicate the same person in each photo.

I'm going to employ Excel to create a spreadsheet to see if there are any striking similarities or differences in the photo, or clumps of similarities. For the spreadsheet, I'm going to eliminate the P/E ratio since it falls within the minimum and maximum range of error for every photo.

[Drew Phipps]

FINAL CONCLUSONS:


The brown line is the minimum value allowing for pixel error, the black line is the maximum value allowing for pixel error. The brightly colored lines are the various ratios.



Visual Conclusions: Unless there is something terribly wrong with my methodology, (or my spreadsheet skills), or the photos are simply too low resolution for a significant biometric comparison, there is some evidence that there is more than one individual here.

If I had to clump the photos in two different piles, it looks to me like photos 1, 2, 3, 9 and 13 are the same individual, and photos 4, 5, 6, 7, 8, 11, and 12 are the same individual.

I realize this is not the same narrative as "Harvey and Lee". It is still my opinion that it is far more likely that the paper trail of discrepancies surrounding Oswald is the result of deliberate tampering with his records (in an attempt to catch a mole, or deceive a communist spy organization). But the biometric discrepancies are starting to make me wonder.

If anyone has any higher resolution digital versions of these shots, that would improve my benchmarks and reduce pixilation errors significantly. Please reply. Also, if you have some other front-facing photos of Oswald you'd like me to compare, please post them below.

[Jim Hargrove]

Thanks, Drew. This was obviously a major undertaking and is very interesting.

From what I understand about facial recognition techniques (which actually amounts to little more than looking at graphics from Facebook and other sources supposedly depicting comparison points on human faces) there are often dozens (possibly even scores or hundreds) of different points of measurement made by these computerized or computer-assisted programs.

For example, in declaring that its new facial comparison algorithm is now "as accurate as the human brain," Facebook offers the following graphic:



This SUGGESTS, at least, that Facebook's approach involves far more comparison points and measurements than the five measurements you used in your study, though I don't mean to disparage your work by suggesting it would be possible for you to carry what you have done much farther in your spare time.

I will bring your work to John A's attention, if he isn't aware of it already, and try to help him find a private company that does this type of work so we can have similar comparisons made. (Assuming this won't be cheap, I'll suggest he start slowly, with a limited number of samples at first.) This is interesting....

[Drew Phipps]

Thank you, Jim. I investigated the published details of Facebook's recognition algorithm ("Facespace"?) and they are talking about projecting and comparing 17 different points on a face. That is far more time and effort than I was willing to invest in the project, especially since I was doing the measurements and calculations myself by hand. I don't entirely trust the "black box" approach: garbage in, garbage out, as they say.

Still, if you want to use Facebook as a platform for evaluating the various claims of identity, might I suggest starting a Facebook page and loading it with a few photos that you label with names. Then drop a new photo in there (without a name) and let Facebook suggest a name. Hardly scientific, but suggestive.

[Jim Hargrove]

After all your work on this, Drew, I'm really embarrassed to ask for some free legal advice, but...

Do you know if the Facebook technique you described above would be in line with FB's Terms of Service? Please don't feel compelled to answer.