Multichannel feature extraction for data-reuploading quantum neural networks
Phys. Rev. Applied 26, 054047 – Published 16 November, 2026
DOI: https://doi.org/10.1103/kprd-ld5d
Abstract
This paper presents a rigorous, quantum-native framework for multichannel image feature extraction utilizing data-reuploading parameterized quantum circuits (PQCs). While PQCs offer significant potential for machine learning tasks in the noisy intermediate-scale quantum era, current architectures are predominantly governed by trial-and-error heuristics rather than mathematical principles. To overcome this lack of rigor, we map complex-unitary operations to real three-dimensional rotations, providing a geometric analysis that links Bloch-sphere trajectories directly to PQC expressivity. Building on this theoretical foundation, we propose a feature extraction mechanism driven by disjoint multipass measurements. Rather than relying on classical convolutional preprocessing, our architecture applies multiple independent measurement passes to individual image tokens. This multipass strategy, combined with data reuploading to capture RGB interdependencies, natively expands the dimensionality of the feature space within a single-qubit framework. Furthermore, to overcome the inherent limits of isolated token processing, we naturally extend our architecture to a multiqubit setting by strictly applying the analytical design rules established at the single-qubit level. Extensive evaluations on the MNIST, ColorMNIST, and CIFAR datasets confirm that this compact, multipass framework achieves the highest reported accuracy among quantum image-processing methods at substantially lower qubit cost. We note that these benchmarks are selected for comparability with existing quantum methods and that classical vision models remain more accurate on them.