Details

This thesis investigates the limitations of classical regularization (L2 and Dropout) in few-shot and noisy image classification, and proposes a lightweight hybrid quantum regularization mechanism combining quantum entanglement and controllable quantum noise. A quantum-classical hybrid framework is built on the MobileNetV1 backbone, validated on a small-sample MNIST dataset under three Gaussian noise intensities (γ=0.1,0.2,0.4). Experimental results show that classical regularization struggles with slow convergence and poor noise robustness under complex conditions. The proposed hybrid quantum regularization significantly accelerates training convergence (77–87 epochs vs. 100 epochs for L2+Dropout). The quantum noise module effectively reduces accuracy decay under medium and strong noise, while the entanglement module stabilizes training. Classical L2+Dropout performs marginally better in generalization error under weak noise, but the proposed hybrid quantum regularization consistently outperforms classical methods in convergence speed and strong-noise robustness, achieving an overall better trade-off. The lightweight framework requires no dedicated quantum hardware and is deployable on conventional computing platforms. This study verifies that quantum characteristics can serve as effective supplements to classical regularization, providing a feasible solution for lightweight models in noisy and data-scarce real-world scenarios.