Cameras can only shrink so much due to physical limitations. A new lensless camera design may change all that.

Researchers are also moving toward a computational approach to focus light on the image sensor, rather than relying on optics at a fixed distance.

Scientists at the Tokyo Institute of Technology have been applying machine learning to prove that a camera doesn’t need a lens. It just needs a new way of seeing the light, much like the Cambrian camera and the sun telescope. This kind of research has been ongoing for many years. However, machine learning is a new way to focus the images and make them more lifelike.

Lensless Cameras Use a computational approach to image processing.

Most designs for a lensless camera use an image sensor that interacts directly with light through a mask. The algorithm then measures that light and reconstructs the image. However, this new approach uses machine learning to analyze every pixel and determine how they interact. Convolutional neural networks (CNN), are then used to reconstruct the image using this data.

While an image may appear, it’s not sharp enough to give any definition or detail without putting a lot of energy and power into solving the computational problem. The image sensor receives only a small amount of light data from the mask, and does not have a lens to focus it. The sensor must therefore work hard to reconstruct the image using that data.

Researchers explain that CNN processes images based on relationships between neighboring pixels. Lensless optics transform local scene information into overlapping, global information on all pixels of the image sensor through a property called “multiplexing”.

A new approach to lensless cameras may solve size limitations in design.

Here is the new approach. Global reasoning can be applied to the entire image sensor using a variety of “Vision Transformers” in order for the light to be identified and analyzed.

“The novelty of this algorithm lies in its structure of multistage transformer blocks with overlapped patchify modules,” the researchers write in a paper published in Physics. This paper was also spotted in DPReview.

This allows it to learn image features in a hierarchical structure. The proposed method is able to address multiplexing and can avoid the limitations of traditional CNN-based deeplearning, which allows for better image reconstruction.

This approach is far more advanced than the CNN process. It relies on neural networks, connected transformers, and yields lower errors. However, it takes less time and resources to analyze and reconstruct the images. Lensless photos can be taken instantly, just like with a traditional camera. Research suggests that additional development can produce higher-quality images with more detail and sharpness.

Lensless cameras can also be made ultra-small, as this has been the goal of lensless camera research. Cameras can be as small or large as they want, provided that they don’t need to follow the rules of physics regarding light bend and distance to create a focused image. In 2013, a lenless camera approach produced a camera image that was one-pixel in width. If this new approach is able to go further than the micron level, cameras might become invisible. The only thing left is to refine the approach.

Prof. Masahiro Yamaguchi, Tokyo Tech, says that the lensless camera can be extremely small without the limitations of a lenses. This could lead to new applications beyond what we have imagined.