Hybrid Classical-Quantum Transfer Learning with Noisy Quantum Circuits

Hybrid classical-quantum architectures have emerged as a practical response to the data and compute demands of modern deep learning, since a pretrained classical backbone can carry the feature extraction while a compact quantum head provides the trainable component. Quantum transfer learning is the most active instance of this idea. However, existing quantum transfer learning pipelines have been evaluated in isolation, typically on a single software framework and without a structured treatment of noise or statistical significance, which makes it difficult to assess how this paradigm contributes over fair classical baselines. A methodological benchmark for quantum transfer learning is proposed in this paper. The benchmark couples a common set of frozen pretrained convolutional backbones to several classical and quantum classification heads, implemented in both PennyLane and Qiskit, so that the contribution of nonlinearity, software framework, and quantum component can be analyzed separately. The benchmark has been applied to heterogeneous image datasets covering medical, biological, industrial, and general-vision domains, and most configurations have been executed in three environments: ideal simulation, noisy emulation calibrated on IBM Heron r2 and real IBM quantum hardware. Complementary analyses isolate the contribution of individual noise channels, examine scalability with qubit count and circuit depth and assess the presence of barren plateaus. Statistical reliability is ensured through multiple random seeds and pairwise Wilcoxon signed-rank tests, providing the first controlled cross-framework assessment of quantum transfer learning under calibrated noise and on real hardware.