QSRNet: Hybrid Quantum-Classical Learning with Variational Circuits for Enhanced Feature Representation in Image Super-Resolution

Abstract

We propose QSRNet, a hybrid quantum-classical framework for single-image super-resolution (SISR) that integrates parameterized quantum circuits (PQCs) within a convolutional encoder-decoder pipeline. QSRNet embeds a quantum variational layer (QVL) comprising 4 qubits and 3 ring-entangled layers into the latent feature space, leveraging quantum superposition and entanglement for expressive, globally-aware feature transformations with only O(nL) trainable parameters. Classical convolutional encoders extract hierarchical spatial features, which the QVL enhances via non-local quantum correlations, followed by lightweight pixelShuffle-based upsampling. Experiments on the datasets Set5, Set14, and DIV2K benchmarks demonstrate that QSRNet achieves 33.05 dB PSNR and 0.912 SSIM on Set5 (×4), outperforming SwinIR (+0.16 dB), HAT, SRFormer, EDSR, and RCAN while using only 9.8M parameters and ≈32.5% faster inference. Comprehensive ablation studies validate the contributions of ring entanglement (+0.25 dB over non-entangled circuits), 4-qubit optimality, shot noise resilience (<0.12 dB drop under depolarizing noise p = 0.01), and composite loss design. These results highlight the potential of quantum-enhanced learning for next-generation image reconstruction.