Speaker
Description
Image reconstruction is a critical problem in industry, especially in certain areas of Optics, such as the Ghost Imaging experiment, [1], [2]. The experiment has many beneficial practical applications such as live cell imaging or remote sensing. The key leverage here lies with its non-local imaging procedure. This allows one to view a quantum image without collapsing its state. The experimental
approach requires twice the number of measurements as opposed to a classical
image, due to the real and complex part of the quantum image. Thus, requires
≈ 2N 2 measurements to reconstruct a N × N image, [3]. The experimental
procedure has challenges in the speed and fidelity of reconstruction. Commonly
used classical reconstruction methods are effective but can be computationally
intensive or struggle to leverage the inherent patterns in natural images.
We have designed a Classical and Quantum algorithm to overcome this intensive computational task. The method we present reconstructs low-sampled
images measured from the Ghost Imaging experiment, using a Classical and
Quantum Convolutional Neural Network (CNN), [4]. Low-sampled images have
a linear representation using the Hadamard transform, where a number of co-
efficients of the linear decomposition are unknown. The CNN’s take the low-
sampled coefficients as inputs and reconstructs the complete set of coefficients.
Instead of directly processing pixel-domain images, our method focuses on re-
constructing missing coefficients in the Hadamard transform domain.
The Quantum CNN model architecture adapts the principles of a Classic
U-Net Convolutional Neural Network. With the use of Variational Circuits, we
will apply the convolutional and pooling layers. Due to quantum properties such
as quantum superposition and quantum entanglement, the model may be able
to exploit more intrinsic patterns and correlation within the Hadamard coeffi-
cient space. We have simulated the Quantum CNN, and seems to show possible
improvement in reconstruction speed and higher fidelity rates, as compared to
its classical counterpart of similar size.
1
This paper will detail the proposed Classical annd Quantum CNN archi-
tecture, the encoding scheme for Hadamard coefficients into quantum states,
the variational quantum layers for feature extraction and upsampling, and the
classical optimization loop. We will present simulation results on the MNIST
data set, and real experimental results from the Wits Structured Light Lab.
Demonstrating the CNN’s ability to reconstruct full Hadamard coefficient sets
from various levels of undersampling, followed by inverse Convolutional Neural
Network to generate high-fidelity pixel-domain images. The findings highlight
the potential of quantum machine learning to significantly advance computa-
tional imaging techniques like Ghost Imaging, paving the way for faster, more
accurate, and quantum-enhanced imaging solutions.
| Presenting Author | Shawal Kassim |
|---|---|
| Institute | University of the Witwatersrand |
