Bringing Grayscale to Ghost Translation
Abstract
In recent decades, physicists developed ghost imaging, which is an alternative technique to the conventional imaging used everywhere by cameras. Ghost imaging was given its name because none of the light that reached the camera-like detector ever interacted with the subject of the image, yet this technique was able to produce an image of that object. This method initially utilized two correlated beams of light; one beam interacted with the object and was collected by a bucket detector without spatial resolution, and the other was simply collected by a spatially-resolved detector. Then, a single-beam method was developed known as computational ghost imaging, where a device modulated the spatial pattern of the only beam. The propagation of the light could be computed, allowing the object to be imaged without a spatially-resolved detector at all. Deep learning techniques were then adopted from computer science to improve computational ghost imaging. In the past few years, researchers have begun utilizing a type of deep learning network called a Transformer network, resulting in a regime known as ghost translation. This regime appears to be robust to noise and could enable major computational shortcuts when compared to ones that utilize other types of neural networks. However, ghost translation has only been developed to work on simple binary images, which do not resemble most applications found in real-life or laboratory settings. I build upon this recent work, exploring the feasibility of extending this regime from binary images to grayscale ones. I take concrete steps toward creating a network that can use such images and identify promising new directions, suggesting that such a network is in the near future. Grayscale images are found in common imaging applications, and they are the step immediately preceding full-color images that are the hallmark of conventional imaging. Hence, improving this regime to give it compatibility with grayscale images opens the door to useful laboratory applications and promotes the discovery of further uses for computational ghost imaging. While a fully-capable grayscale Transformer network is not yet here, I bring it several steps closer in this work.
Citation
Bottorff, Zachary David (2023). Bringing Grayscale to Ghost Translation. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /199652.