- MIT researchers use deep neural networks to reveal invisible objects in complete darkness.
- It can image biological cells and tissue without exposing them to harmful rays and intense light.
Most imaging systems only yield distorted or partial data about the object being imaged. This happens mainly due to lack of phase information, loss of spatial frequencies, noise in the illumination, and unknown scatterers in the optical train.
Over the last few couples of years, a machine learning technique called deep neural network has attracted much attention in the computational imaging field. It has been proven an efficient solver in a variety of applications, including ghost imaging, adaptive optics, adaptive illumination microscopy, phase retrieval, optical tomography, and undersampled imaging.
For the first, a team of MIT researchers has used deep neural networks to solve a coherent phase retrieval problem affected by high noise at different levels. In simple language, they have discovered a method to reveal invisible objects in complete darkness.
This is much different than existing AI-based ‘night-mode’ technique found in Google’s Pixel 3 smartphone, which can capture several noisy images and produce clear photos, but it requires some light to begin with. MIT’s technique, on the other hand, works in a fully dark room. It requires only one photon per pixel.
How They Did This?
The researchers first captured pictures of target objects in almost pitch-black conditions. They then recreated transparent objects from these pictures. To do this they used deep neural networks, which is trained to recognize over 10,000 transparent glass-like etching from dark, grainy pictures that are invisible to the human eye.
The pictures themselves, captured in a dark room, seemed like the static noise you could see on a television. The neural networks are trained on these images along with the corresponding patterns beneath the visual noise.
Gradually, the network learned to make sense out of the visual noise. It eventually generated blurry pictures. To make these pictures more clear, the team added a layer that can focus the output.
Courtesy of researchers
In the above image, you can see the dark image (top left), which is generated from a transparent etching (extreme right). The researchers used physics-algorithm based on the behavior of light to recreate the object (top right). The machine learning technique created quite a blurry image (bottom left). They combined both physical-algorithm and machine learning technique to reconstruct the most accurate image (bottom right) of an actual object/scene.
How Is It Useful?
The AI could be used to illuminate transparent features like biological cells and tissues, in pictures captured with very low light. Cells can be easily burned or damaged if they are exposed to intense light, and then there will be nothing left to image. Also, when patients are exposed to X-rays, they have chances of developing cancer.
This study could help in such cases: researchers have made it possible to get the same image quality by exposing cells and tissues to fewer photons. This significantly decreases the damage to biological specimens when sampling them. Moreover, the technology could offer a range of potentially useful applications in the field of astronomical imaging.