Benchmarking of Quantum Image Dataset Transform Pre-processing Across Hybrid Quantum Architectures
Published in Workshop on Advancing Hybrid Quantum-Classical Computing Through Shared Challenges, IEEE International Conference on Quantum Computing and Engineering (QCE), 2026
Quantum machine learning (QML) enables hybrid architectures for image classification on Noisy Intermediate-Scale Quantum (NISQ) hardware. However, the impact of quantumnative data pre-processing remains underexplored. In this paper, we conduct an initial study to benchmark hybrid quantum models with and without the Quantum Image Dataset Transform (QIDT). The framework trains matched pipelines on raw and QIDTtransformed images, with data transformation as the sole variable, and measures classification accuracy, throughput, and training energy across the MNIST, Fashion-MNIST, and MedMNIST datasets same as the existing works. Results show that on PneumoniaMNIST, models retain accuracy under QIDT, with paired differences remaining within the run-to-run spread across five seeds. On MedMNIST, QIDT’s effect varies: though having high throughput and low energy consumption it reduces accuracy on OrganAMNIST but improves accuracy and macro F1 on classimbalanced subsets like DermaMNIST and BreastMNIST. QIDT’s 16-dimensional encoding front-end for QCNNs falls between PCA and autoencoders in reduction performance on MNIST and Fashion-MNIST. Performance analysis reveals that QIDT occurs only once before training, but per-patch quanvolution drastically reduces throughput to 2.3 samples/s and increases training energy (12×) compared with the DH-QCNN. Overall, the impact of quantum data pre-processing depends on architecture and dataset, with quantum pipeline placement dominating practical feasibility.