Deep Neural Network AI Reconstructs Mysterious Image Hidden in Picasso Painting
As artificial intelligence continues to develop apace, it’s clear that AI will in all likelihood make huge gains in a wide range of fields, extending from making medical diagnoses and financial predictions, to automated fact-checking. What’s less apparent is what kind of role AI will play in more creative industries, as it’s recently done in generating surprisingly compelling works of art, literature and music.
While many might disagree about the artistic ability of machines — no matter how intelligent — AI could still make significant contributions in this realm. One excellent example comes from a team of University College London researchers, who recently used AI to reconstruct a never-before-seen painting of a nude woman by famed Spanish painter Pablo Picasso, which was apparently hidden under The Old Guitarist. This celebrated work hails from Picasso’s so-called “Blue Period” of the early 1900s, a time of financial scarcity and psychological depression during which the master artist painted almost exclusively in somber hues of blue.
Using a computer vision technique known as neural style transfer (NST), the researchers were able to reconstitute the ghostly image of a woman hidden underneath, which was later painted over by Picasso to create The Old Guitarist (reusing canvases being one way that impoverished artists could save money). By employing the neural style transfer technique, the team was able to recover the image without having to damage the artwork itself, with better results than conventional methods like using x-ray imaging.
La Femme Perdue
As the team outlines in their paper, neural style transfer refers to a class of algorithms that manipulate digital images and videos so that the resulting output adopts the visual style of a reference image. First developed in 2015, NST algorithms use what is known as convolutional neural networks, a type of multilayered, deep learning AI that’s often used in image and video recognition and classification tasks, medical image analysis and recommendation systems.
Neural style transfer works by having the user provide a “content” image and a “style reference” image — such as an artwork by a famous artist — plus an “input” image that needs to be styled according to the visual aesthetic of the “style reference” image. These are blended so that the “input” image is altered to look like the “content” image, but stylized to look like the “style reference” image.
In the case of this study, the team already had the faint, x-rayed outlines of the woman underneath the old guitarist, so the next step was to manually edit out any features of the topmost painting that were unlikely to be in the original painting underneath. With The Old Guitarist set as their “content” image, the team then took Picasso’s painting La Vie as their “style reference” image. By applying neural style transfer, the team was then able to reconstruct the hidden painting of La Femme Perdue (“the lost woman”) — in colorful brushstrokes that Picasso might have painted himself. Intriguingly, these aesthetic attributes can also be tweaked via what is known as error functions and optimization algorithms, so that an image can take on as much or as little of a particular style.
The Emergence of Artificial Creativity?
More broadly, such methods will likely change the way art historians will operate in the future, says PhD student George Cann, one of the paper’s co-authors and co-founder of AI art collective Oxia Palus: “I think the emergence of AI in art will significantly expand our creative horizon. We’re at the creative shoreline taking our first steps on land. In the field of art history, AI will help by making it easier and cheaper to reveal what lies beneath historical artwork. There are potentially thousands of underpaintings and underdrawings, each with their own story, that are hidden from the world. This hidden art may hold important information about how art has developed throughout history.”
Understandably, there is still plenty of skepticism about whether machines can be truly creative. But does artificial creativity have to be a faithful copy of the human version of creativity? Might it be possible that machines might develop their own unique sense of innovation? If nothing else, integrating AI into humanity’s creative processes might prompt more people to shift their presumptions about AI and creativity.
“We believe that AI art, broadly, challenges our perception of creativity — a thing that many believe to be an inherently sentient property,” said paper co-author Anthony Bourached, who is yet another founding member of Oxia Palus. “We call this belief the ’empathy paradox.’ In a more sociological vein, we also believe there is a very positive message that human-AI collaboration broadens the beachhead into creative domains, rather than narrowing it. In the long term, AI does not ‘replace people,’ but instead makes us more fulfilled in the same way that a sturdy set of boots broadens the spectrum of turf on which we may march.”
Read the paper here.