QUSL: Quantum Unsupervised Image Similarity Learning with Enhanced Performance

This is the authors' abstract. We don't add key points for this paper.

Leveraging quantum advantages to enhance machine learning capabilities has become a primary focus of research, particularly for complex tasks such as image similarity detection. To fully exploit the potential of quantum computing, it is essential to design quantum circuits tailored to the specific characteristics of the task at hand. In response to this challenge, we propose a novel quantum unsupervised similarity learning method, QUSL. Building upon the foundation of similarity detection triplets and generating positive samples through perturbations of anchor images, QUSL operates independently of classical oracles. By leveraging the performance of triplets and the characteristics of quantum circuits, QUSL systematically explores high-performance quantum circuit architectures customized for dataset features using metaheuristic algorithms, thereby achieving efficient quantum feature extraction with reduced circuit costs. Comprehensive experiments demonstrate QUSL's remarkable performance compared to state-of-the-art quantum methods. QUSL achieves reductions exceeding 50% in critical quantum resource utilization while also realizing an enhancement of approximately 20% in similarity detection correlation across the DISC21, COCO, and landscape datasets. This enables efficient quantum similarity modeling for large-scale unlabeled image data with reduced quantum resource utilization.