Overview of Quantum Computing in Image Processing

How can classic images be converted into quantum information – and what are the benefits for image processing? At Fraunhofer ITWM, we are researching various approaches to quantum image processing, from efficient encoding to classification and segmentation to edge detection.

Our work shows that the current state of quantum hardware already enables promising approaches for future image processing tasks using hybrid quantum algorithms. This page provides an overview of the applications and current research:

Quantum Image Encoding

Quantum Image Classification with Quantum Transfer Learning

Quantum Image Classification with VQLS und SVM

Quantum Image Edge Detection

Quantum Image Segmentation with Hamiltonian and Q-Seg

Quantum Fourier Transform for Angle Estimation

Encoding classical images into quantum states is a fundamental step in quantum image processing. Various encoding methods exist, such as basis, amplitude, and phase encoding. In our research, we focused on enhancing the Flexible Representation of Quantum Images (FRQI) encoding by reducing the number of controlled operations. This optimization significantly decreases error rates and improves the feasibility of running quantum image processing tasks on current Noisy Intermediate-Scale Quantum (NISQ) hardware.

Our findings show that while large-scale image encoding remains challenging, improved circuit efficiency and noise mitigation techniques make quantum image representation a promising direction for future applications.

Quantum transfer learning combines classical deep learning with quantum computation to enhance image classification. In our study, we applied this approach to detect fine cracks in concrete images (see HairWidthCracks dataset). By leveraging a pre-trained classical model for feature extraction and a variational quantum circuit for classification, we demonstrated the potential of hybrid quantum-classical models.

While quantum transfer learning helps bypass limitations in qubit count and error rates, training remains constrained by the need for repeated circuit evaluations due to the lack of quantum backpropagation. Despite these challenges, our results indicate that quantum-enhanced classification can be successfully implemented on current quantum hardware.

Variational Quantum Linear Solvers (VQLS) present an alternative approach to quantum image classification by solving linear systems using variational circuits. Our research integrates VQLS into a hybrid quantum-classical classification framework, optimizing quantum parameters through iterative feedback loops.

The primary challenge is the non-convex loss landscape, which requires extensive circuit evaluations to refine the model. While VQLS-based classification holds theoretical advantages in efficiency, current quantum hardware introduces noise and stability issues, necessitating algorithmic modifications for practical deployment.

Quantum computing provides a novel way to detect edges in images, leveraging the efficiency of quantum circuits. In our work, we adapted Tacchino’s quantum artificial neuron model to develop a quantum edge detection algorithm. This approach processes image data using quantum parallelism, reducing the computational complexity compared to classical methods like Sobel filtering.

However, due to hardware noise and gate limitations, we designed multiple algorithm variations to optimize execution on NISQ devices. Our results demonstrate that quantum-based edge detection can handle larger images than previously feasible, marking an important step toward practical quantum image processing.

Quantum image segmentation leverages quantum computing principles to efficiently partition images into meaningful regions. In our research, we explored two quantum-based approaches: a Hamiltonian-based method and Q-Seg, a quantum annealing algorithm. The Hamiltonian method embeds image data into a quantum system where disorder effects naturally highlight cracks and edges, making it highly effective for detecting complex patterns. Q-Seg, on the other hand, formulates segmentation as a combinatorial optimization problem and solves it using quantum annealing, demonstrating strong performance in unsupervised learning tasks.

Our study, benchmarked against classical methods like Gaussian Mixture Models and deep-learning-based U-Net, reveals that while quantum approaches are still constrained by current hardware limitations, they offer promising results, particularly in detecting fine-grained structures like cracks in concrete surfaces

The Quantum Fourier Transform (QFT) is a powerful tool for analyzing frequency components in quantum image processing. We applied QFT to estimate object orientations in images, comparing its performance with the classical Fast Fourier Transform (FFT). While QFT theoretically offers exponential speedup, its practical implementation is hindered by the high number of entangling gates required, leading to increased noise and decoherence.

Our experiments show that QFT-based angle estimation can achieve results comparable to classical FFT on simulators, but real hardware limitations must be addressed for practical applications.