Somerton Man Image Restoration
Neural-Network assisted reconstruction of bust and post-autopsy photo for Tamam Shud case
WARNING: The end of this article contains images of a deceased man. There will be a second warning at the bottom of this article.
Using Photoshop and the neural-net tool Artbreeder I have created a digital reconstruction of Somerton man as he might have appeared in the weeks before his death in 1948.
RE IMAGE RIGHTS: You may republish the reconstructions as long as the watermark is not removed or obscured. CC-BY-ND-4.0
The image restoration was guided by research and notes by University of Adelaide Professor Derek Abbott introduced to me by cold case forensic genealogist Colleen Fitzpatrick.
Notes on appearance via Derek Abbott (via Inquest and Police files)
- His legs were suntanned. So he had the skin type that can tan. So if you look at the skin tone of red heads, his tone would probably not be that pale. Red heads don’t usually tan that well….they burn.
- His canines are next to his middle teeth. (He has no lateral incisors._
- He wore his hair combed back with no parting.
- His hair colour is described as “mousey with flecks of red and greying at the sides”
- His eyes are described as “grey blue”
- Neck folds are an effect of the autopsy
- The way his shoulders are shrugged up showing no deltoid muscles is probably because he was pulled by the shoulders to get into the freezer.
- Bust was made 6 months after death, so the face has flattened. Back of head has flattened too.
- Photos were taken after the autopsy 1–2 days after death.
Other notes via Derek Abbott’s University of Adelaide lecture:
- Mole above upper left lip
- Large upper hollow in left ear (little known about right ear)
- Attached lobes
Personal Notes / Process
An unbiased artistic interpretation is a challenge. The peculiar nature of the case fueled my interest and I was made aware of a possible son of the Somerton Man. I did not use the presumed son in my reconstruction but having seen the photo is can cause an unconscious bias in the decision making process.
- Neural net was trained on previous reconstruction photo but little was used in the mix. Maybe 10%.
- Roughly 20% bust
- 70% the autopsy photo.
- No use of Robin in ‘Tasmania photo’. (Not even for teeth. I used another photo for reference.)
Below are the two main image restorations that when combined and transformed, to add a smile and open mouth to show teeth, resulted in the final image.
One is a photo taken 1–2 days after death. The other was a plaster cast taken 6 months after death.
ABOUT THE AUTHOR
Daniel Voshart is a designer from Toronto, Canada who dabbles in video analysis and 3D scene reconstruction. He is also testing the uses of machine learning in forensics.
ABOUT FORENSIC VR
Forensic VR is the umbrella brand for Daniel Voshart’s forensic-related work.