Hybrid Quantum--Classical Pauli-Correlator Architecture for Image Reconstruction
Organizations: School of Basic Sciences, Department of Physics, Indian Institute of Technology Bhubaneswar, Odisha, India. · Centre for Quantum Engineering, Research and Education, TCG CREST, Kolkata, India.
Abstract
We introduce the Pauli-Correlator Autoencoder (PCAE), a hybrid quantum--classical architecture for image reconstruction. PCAE splits an image into patches and compresses each patch as a matrix product state (MPS). A small quantum circuit turns each compressed patch into a few measured numbers, and a classical neural network rebuilds the patch from these numbers and the patch position. A learnable energy term, taken from physics, is added to training and tracked as training proceeds. Two versions place the six qubits on a line (PCAE1D) or on a grid (PCAE2D). An averaging step makes the PCAE2D features follow rotations and reflections of the image exactly. Trained on one image at a time, PCAE2D rebuilds images from six public datasets with a peak signal-to-noise ratio of 25.8 to 41.6 decibels. On CIFAR-10 images not used in training, a formula-based stand-in for the PCAE2D features performs within 1 decibel of several simpler classical methods. Five trained models also run on real IBM quantum computers without retraining, and rebuild images at 25.90 to 31.52 decibels. A short argument and a synthetic test show that, with a linear final layer, measuring qubits in pairs is needed for targets built from products of inputs. These results show that PCAE works as a hybrid model that rebuilds varied images and runs on today's quantum hardware at the six-qubit scale.
Figures & tables
| PCAE1D | PCAE2D | |
| Benchmark | Monalisa | multi-dataset |
| Image size | (resized) | |
| Patch size | ||
| MPS sites | 12 | 6 |
| Local expectations | 24 | 12 |
| Nearest-neighbour | 11 | 5 |
| Variant | Topology | Latent dim. | Symmetry |
|---|---|---|---|
| PCAE1D | 6-qubit chain | 27-D | Unpooled |
| PCAE2D | grid | 19-D | Unpooled |
| -pooled PCAE2D | grid, pooled | 19-D | -equivariant (exact) |
| PCAE2D | Position-only ablation | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Dataset | images | PSNR (dB) | SSIM | PSNR (dB) | SSIM | ||||
| CIFAR-10 | 3 | 37.75 | 2.38 | 0.9986 | 0.0003 | 42.98 | 8.54 | 0.9991 | 0.0008 |
| MNIST | 3 | 25.77 | 2.23 | 0.9783 | 0.0116 | 25.95 | 2.31 | 0.9796 | 0.0101 |
| Fashion-MNIST | 3 | 34.55 | 3.05 | 0.9980 | 0.0006 | 33.70 | 0.49 | 0.9968 | 0.0026 |
| PathMNIST | 3 | 36.86 | 2.11 | 0.9118 | 0.0627 | 35.49 | 1.92 | 0.8904 | 0.0655 |
| BloodMNIST | 2 | 36.57 | 2.36 | 0.9922 | 0.0063 | 34.67 | 3.98 | 0.9833 | 0.0151 |
| Control | PSNR (dB) | 90% CI (dB) | TOST | Holm | Verdict |
|---|---|---|---|---|---|
| Raw/local linear | Equivalent | ||||
| No entanglement | Equivalent | ||||
| -only observables | Equivalent | ||||
| Quadratic classical | Equivalent | ||||
| Shuffled pair observables | Equivalent | ||||
| Strict classical random features | PCAE2D better |
| Image | Patches | Sim. PSNR (dB) | HW PSNR (dB) |
|---|---|---|---|
| Monalisa (PCAE1D) | 16 | 37.99 | 25.90 |
| CIFAR cat (PCAE2D) | 16 | 44.85 | 31.52 |
| CIFAR ship, hydrofoil (PCAE2D) | 16 | 36.56 | 26.44 |
| CIFAR ship, sea boat (PCAE2D) | 16 | 42.57 | 31.20 |
| CIFAR airplane (PCAE2D) | 16 | 46.79 | 29.10 |
| CIFAR mean std (4 images) | — |
| Image | Backend | Shots | PSNR (dB) |
|---|---|---|---|
| Monalisa | ibm_fez (original) | 256 | 25.90 |
| Monalisa | ibm_marrakesh | 256 | 25.94 |
| Monalisa | ibm_kingston | 1024 | 26.10 |
| CIFAR cat | ibm_fez (original) | 256 | 31.52 |
| CIFAR cat | ibm_kingston | 256 | 28.93 |
| CIFAR cat | ibm_kingston | 1024 | 31.24 |
| Model | Mean error | Std. dev. |
|---|---|---|
| -pooled PCAE2D | ||
| No positional encoding | ||
| PCAE2D positional | ||
| Local/raw |
| Model | Mean MSE | Mean PSNR (dB) | Mean |
|---|---|---|---|
| PCAE2D pair observables | 32.77 | ||
| Freeze quantum | 27.34 | ||
| No pair observables | 15.03 | ||
| Raw-linear classical | 15.00 | ||
| Parameter-matched classical | 14.82 | ||
| Random VQC | 14.25 |
| Experiment | PSNR (dB) | Interpretation |
|---|---|---|
| Legacy baseline, signed energy | 38.63 | Standard energy-regularised run |
| No energy loss | 44.56 | Removing energy improves PSNR |
| Contrastive Hamiltonian core | 42.92 | Stronger objective helps modestly vs. baseline, still below no-energy-loss |
| Axis | Conv. AE baseline | GAN baseline | Register-based QAE | Quantum-kernel vision | TN-circuit classifier | PCAE (ours) |
|---|---|---|---|---|---|---|
| Latent representation | Continuous | Continuous | Quantum register | Scalar kernel value | MPS/circuit state | Pauli-correlator vector |
| Spatial bias | Convolutional | Convolutional | Encoding-dependent | Encoding-dependent | MPS chain ordering | 2D Fourier encoding |
| Auxiliary objective | Bottleneck / reconstruction | Adversarial loss | Fidelity / reconstruction | None (kernel value only) | Classification loss | Hamiltonian energy term |
| Enforced group symmetry | Not in baseline | Not in baseline | Not in reference design | Not in reference design | Not in reference design | observable pooling |
| Evaluated role | Reconstruction | Reconstruction | Compression | Classification | Classification | Reconstruction + diagnostics |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Experiment | Model | Patches | Steps | Learning rate | Seeds |
|---|---|---|---|---|---|
| Monalisa hardware checkpoint a | PCAE1D | 16, | 200 | 0.003 | 42 |
| 240-step ablation | PCAE1D | 16, | 240 | 0.003 | 0, 1, 2 b |
| Six-dataset fits | PCAE2D | 16, | 100 | 0.004 | not recorded |
| CIFAR-10 hardware fits | PCAE2D | 16, | 200 | 0.004 | not recorded |
| Topology comparison | 1D / 2D | 49, , stride 4 | 90 | 0.003 / 0.004 | 0, 1, 2 |
| Held-out comparison | surrogate | 16, | n/a | n/a (ridge) | 0–4 c |
| Variant | PSNR (dB) | SSIM | Variant | PSNR (dB) | SSIM |
|---|---|---|---|---|---|
| Baseline (quantum) | Classical-replacement | ||||
| Freeze-quantum | Classical-matched | ||||
| No-entanglement | Random-VQC | ||||
| No-MPS | Freeze-classical | ||||
| No-energy-loss |
| Dataset | Best local/raw | PCAE2D surrogate | Higher mean |
|---|---|---|---|
| CIFAR-10 | Local/raw | ||
| MNIST | No-entanglement/local | ||
| Fashion-MNIST | Local/raw | ||
| PathMNIST | Local/raw | ||
| BloodMNIST | Local/raw | ||
| PneumoniaMNIST | Local/raw |
Image Reconstruction
Image reconstruction recovers images, volumes, or scenes from incomplete, noisy, indirect, or compressed measurements using physical models and learned priors.
See all 312 papers