Shivanshu Siyanwal

Journal article

Quantum image representation-based quantum neural networks for binary classification

A. Tiwari*, S. Siyanwal*, S. Sakhuja, S. Kumar, P. Tyagi

The Journal of Supercomputing, 2026 · published

Overview

A controlled comparison of four quantum image representations (NASS, QBIR, FTQR, FRQCI) inside one variational quantum neural network framework for image classification on near-term (NISQ) devices, tested on binary MNIST (0 vs 1) and a three-class FashionMNIST subset at resolutions from 2×2 to 16×16, under both noiseless simulation and a calibrated noise model of the IBM Heron r3 processor. Two reference models, a quanvolutional network (QCNN) and a hybrid CNN–QNN (HCQNN), place the encodings against established baselines.

Problem / motivation

Quantum image classification depends heavily on how classical pixel data is encoded into quantum states, but encodings are often chosen by convention rather than by measured trade-off.

Methods

Four quantum image representations — Normal Arbitrary Superposition State (NASS), Quantum Block Image Representation (QBIR), Fourier-based Threshold Quantum Representation (FTQR) and Flexible Representation for Quantum Color Images (FRQCI) — are compared within one variational QNN framework. The main protocol runs the four encodings under a ring-entangled ansatz from 2×2 to 8×8; an ablation repeats the comparison with a hardware-efficient two-design ansatz and a strongly entangling ansatz, at all four resolutions on both datasets. Resource cost is measured as circuit depth, gate count and qubit count, and trainability through per-epoch gradient variance, L2 gradient norm, mean signed gradient and parameter-change magnitude. Every result is a mean ± σ over 20 random seeds.

Results / outcomes

HCQNN was the strongest overall, with test accuracy up to 0.9756 ± 0.0076 (2×2, noiseless, two-design). Among the fully quantum models, QBIR gave the best balance of accuracy, noise robustness and depth: 0.9396 ± 0.0140 (2×2, noiseless, two-design), and the strongest fully quantum encoding at every resolution under Heron r3 noise. NASS was competitive at low resolution (up to 0.8928 ± 0.0196) but fell to 0.691 at 4×4 under the ring protocol. FTQR and FRQCI did not converge reliably, with test accuracy near chance (≈ 0.50–0.61), consistent with barren plateaus. At 16×16, QBIR needs depth 5,612, 9,701 gates and 16 qubits; FRQCI needs depth 125,299, 164,353 gates and 11 qubits.

My role

I conceived this study: the question of how the choice of quantum image encoding shapes the trainability, noise robustness and cost of variational QNNs. I designed the experimental protocol, selected and implemented the four encodings (NASS, QBIR, FTQR, FRQCI) in a single QNN framework, and carried out the comparative analysis, the visualisations and the organisation of results. During revision, in response to referee and editor feedback from The Journal of Supercomputing, I also carried out the state-of-the-art comparison, the gradient (trainability) analysis and the mathematical modelling.

ProblemApproachResultsStatus
Schematic: structure of the work (from the write-up's sections)

Project write-up with figures →