A camera equipped with 3D-printed plastic lenses designed by artificial intelligence can take photographs that capture only certain objects and ignore everything else in the frame. The selective lenses use no power and instead rely on complex patterns printed into the plastic to diffract light in a different way.
at the University of California, Los Angeles, and his colleagues trained a deep-learning AI model to recognise images of numbers from a huge online repository of handwritten digits inside a simulation. Some numbers were designated as target ones that would appear in images, some as objects to ignore.
The AI was shown thousands of these images, one at a time, and told when those that were supposed to reach the camera鈥檚 sensor did and didn鈥檛 pass through a series of three lenses, as well as when numbers that weren鈥檛 supposed to reach the sensor did. As these results were given to the AI, it tweaked its lens design to better achieve the desired effect.
Advertisement
Once the model converged on a working solution within the simulation, the lenses were 3D printed with clear plastic, and tested in the laboratory on more images of numbers from the same repository 鈥� ones that hadn鈥檛 been used in the training phase. Each target image was placed in turn between the AI-designed lenses and a terahertz radiation source that was used instead of visible light, with a sensor placed at the other end of the series of lenses.
The researchers found that the AI-designed camera successfully allowed the target numbers to be photographed, while diffracting away light relating to anything else so that it didn鈥檛 appear in the final image. Other tests on images of items of clothing such as trainers and trousers were also successful.
Ozcan says the AI learns to create lenses that would be impossible to design mathematically from scratch. 鈥淏asically, it鈥檚 a puzzle,鈥� he says. 鈥淢athematically, I don鈥檛 know what trousers are, right? I cannot define mathematically what trousers are. And that鈥檚 where deep learning comes into place. That鈥檚 the power of deep learning-based design, because mathematically, what this does is difficult to explain.鈥�
The AI creates designs for the lenses that allow certain shapes from objects to pass through, but diffracts others so that the photons of light coming from them are scattered sideways and never reach the camera鈥檚 sensor. This means that the processing is done physically by the lenses and no computer is needed, and the unwanted objects are never captured digitally 鈥� rather than being captured and then edited out.
The camera听uses terahertz radiation because it is better suited to the 3D-printed lenses that are quicker and cheaper to prototype in a laboratory than ones that worked with visible light would be, but Ozcan says that the visible spectrum 鈥� and even multicoloured images 鈥� would be possible with better quality and finer lenses.
Ultimately, Ozcan says that such lenses could be as thin as 1.2 millimetres and just 2 millimetres square, and be easily fitted to even tiny cameras such as those in smartphones.
The computation is done physically so there is no additional power requirement above a standard camera, but because the lenses are physical, not digital, they can鈥檛 be altered once manufactured. This means they will only ever be suitable for one application and one set of target shapes.
So although the AI lenses could theoretically be capable of learning advanced tasks such as only photographing your friends and family, and ignoring other people in frame, they are more likely to be used in industrial and security situations where images would have a more consistent shape.
They could be trained to only see broken components in a factory to help with quality control or to look for specific weapons in a security application, for instance.
eLight
