Quantum-Classical Hybrid K-means Image Segmentation: Fusion of CUDA Acceleration and Quantum-Inspired Algorithms

This project is the core practical part of the undergraduate thesis at Babeș-Bolyai University (BBU). It implements real-time image segmentation using C++ and CUDA, innovatively fusing classical GPU acceleration with quantum-inspired distance measurement methods to explore the application path of quantum machine learning (QML) in the field of high-performance computing. The project source code is available on GitHub (link: https://github.com/sebsop/kmeans-thesis-segmentation) and was released on June 6, 2026.

• The author got involved in quantum computing in the second year of undergraduate studies and completed several related projects; integrating quantum principles into the graduation thesis was a personal goal.
• The project is an intersection of high-performance computing, artificial intelligence, and quantum mechanics, laying the foundation for future QML research.
• Evolved from the 'Real-time Parallel K-means Image Segmentation' project in a previous parallel and distributed programming course, shifting from distributed CPU parallelism (MPI/OpenMP) to single-node GPU acceleration and quantum-inspired similarity measurement.

Limitations of Traditional Euclidean Distance

• Assumes spherical data distribution, poor performance on non-convex clustering; sensitive to high-dimensional data; unable to capture complex non-linear relationships.

Quantum-Inspired Phase Estimation Measurement

• Simulated Swap-Test interference approximation: Simulates quantum interference effects via GPU, calculates phase overlap of vectors in Hilbert space, and captures quantum entanglement features that traditional Euclidean distance cannot express.
• Swap-Test principle: A basic operation in quantum computing to measure the similarity of two quantum states, estimating the inner product through interference patterns.

Dual-Engine Hot Switching

• Classical CUDA Engine: Optimizes traditional K-means, Euclidean distance clustering, CUDA kernel space preprocessing, K-means++ initialization.
• Quantum-Inspired Engine: Simulates Swap-Test interference, Hilbert space phase overlap calculation, non-Euclidean similarity measurement, shares CUDA parallel architecture.

CUDA Kernel Optimization

• Spatial preprocessing: Shared memory reduces global access, parallel pixel preprocessing, supports color space conversion.
• K-means++ initialization: Parallel centroid selection, optimized distance matrix calculation.
• Pixel assignment: Large-scale parallel distance calculation, shared memory caching of centroids, thread-safe atomic operations.

Temporal Coherence Optimization

• Centroid memory mechanism: Reuses previous frame centroids in stable scenes, recalculates when scenes change, improving throughput for high-frame-rate video processing.

The project integrates real-time metric calculation and supports multiple clustering quality metrics (all CUDA-optimized without affecting frame rate):

• Approximate Silhouette Score: Measures the comparison between a sample's similarity to its own cluster and to other clusters.
• Davies-Bouldin Index: Ratio of intra-cluster to inter-cluster distances; the smaller the value, the better the quality.
• Within-Cluster Sum of Squares (WCSS): Classical K-means objective function, measures cluster compactness. These metrics support fair comparison between classical and quantum-inspired methods.

Design Patterns

• Factory pattern: Dynamically instantiates execution engines and initializers, enabling transparent switching.
• Observer pattern: Decouples CUDA processing loops from UI rendering threads to avoid lag.

Code Quality

• Doxygen documentation, C++17 standard (smart pointers, RAII), GoogleTest unit tests, clang-format/tidy to ensure consistent style.

Project Structure

Mainly includes modules such as backend (CUDA interface), clustering (core logic: engines, initializers, preprocessing), common (shared configuration), and tests (unit tests).

Dear ImGui UI

• Dynamic parameter control (K value, number of iterations, etc.), real-time telemetry (FPS, GPU utilization, clustering metrics), side-by-side visual comparison (original video, classical/quantum segmentation results), one-click engine switching.

Research Value

• Exploring quantum-classical hybrid computing: Using classical hardware simulation to verify the potential of quantum algorithms when quantum hardware is not mature.
• Application scenarios: Medical image segmentation, video surveillance foreground separation, low-latency real-time image processing.

Future Directions

• Integration with real quantum hardware;
• Exploring the application of Variational Quantum Algorithms (VQA) in clustering;
• Multi-GPU distributed expansion;
• Fusing CNN feature extraction with quantum-inspired clustering.