Source-linked AI summary
Facial classification Using Hybrid Quantum Machine Learning
Roshan Babu Bandlapalli, Srinivas V Katakam, Jitendra Chougala, Ravi Kumar Kappagantu, Jayasri Dontabhaktuni
TL;DR
Hybrid quantum methods have received limited study for resource-constrained facial biometrics. The paper combines classical preprocessing and dimensionality reduction with an eight-qubit variational classifier, reporting 97% binary accuracy and training under one hour on standard CPU hardware. The authors conclude that such systems can offer efficient alternatives for specific biometric tasks where high-performance GPU infrastructure is unavailable.
Problem
Hybrid quantum methods for facial biometrics have limited evidence on realistic-scale accuracy, deployment conditions, and equivalent-hardware comparisons.
Source: §1 · pp. 2, 4Method
The pipeline applies classical preprocessing and dimensionality reduction before an eight-qubit variational quantum classifier models nonlinear decision boundaries.
Source: §1 · pp. 2–3Results
97% binary classification accuracy was achieved, with training under one hour on standard CPU hardware, versus 85.1% accuracy and significantly longer training for the classical baseline.
Source: §5 · p. 14Takeaways & Limitations
The results indicate that hybrid quantum–classical systems can offer efficient alternatives for specific biometric tasks where high-performance GPU infrastructure is unavailable.
Source: §5 · p. 15The system classifies face versus non-face, while real-world authentication requires identifying specific individuals among many.
Source: §4 · p. 14
Abstract
from arXiv · showhide
Hybrid quantum methods have received limited study for resource-constrained facial biometrics. We present a hybrid quantum-classical facial recognition pipeline designed to run on standard computing hardware. Images undergo gamma correction, contrast enhancement, and principal component analysis before their features are encoded into an eight-qubit variational quantum classifier. Classical image matching then performs recognition. In experiments with 50,000 images, comprising 25,000 faces from CelebA and 25,000 non-face images from CIFAR-10, the method outperformed the reported CPU-trained FaceNet baseline in accuracy and training efficiency. The pipeline was also evaluated on GPU and quantum hardware. In an attendance monitoring deployment at Mahindra University in collaboration with Lloyds Technology Centre, CPU inference took 0.2 to 0.5 seconds per person, and the system remained robust to the use of spectacles. These findings support the feasibility of deploying hybrid quantum methods for facial recognition on existing CPU hardware.
1 Introduction
Prior hybrid quantum facial-biometrics work leaves gaps in realistic-scale accuracy, practical deployment, and equivalent-hardware comparisons. This study proposes a resource-conscious classical–quantum pipeline that combines classical feature processing with a variational quantum classifier.
1 Introduction
- The pipeline applies illumination and contrast preprocessing and PCA before angle-encoding compact features into an entangling variational quantum classifier.Classical processing handles high-dimensional features, while the quantum component learns nonlinear decision boundaries.Source: §1 · p. 2
- The authors position the framework for resource-constrained settings by reducing parameter count and feature dimensionality before quantum processing.They describe the approach as a potential fit for organizational and edge environments with limited access to high-end GPUs or large annotated datasets.Source: §1 · p. 2
- The study asks whether a hybrid system can match or exceed classical baselines on a substantial facial task and work reliably in real-world deployment.These are the two stated research questions.Source: §1 · p. 2
- The authors report that earlier quantum support vector machine implementations were computationally intensive and difficult to scale in their experiments.They describe extended training and suboptimal classification accuracy before exploring variational classifiers.Source: §1 · p. 2
- Prior biometric studies include quantum facial recognition at 95% accuracy on 400 Olivetti images, while practical deployment was not addressed.The cited work used four qubits and 40 subjects.Source: §1 · p. 4
- The study identifies limited realistic-benchmark accuracy, practical deployment evidence, and equivalent-hardware comparisons as gaps in hybrid facial biometrics.It frames its response as 50,000-image trials, 97% accuracy, performance analysis, and deployment validation.Source: §1 · p. 4
2 METHODOLOGY
The methodology combines lightweight classical face preprocessing and dimensionality reduction with an eight-qubit variational classifier. Its selected angle encoding and three-layer ansatz reflect reported trade-offs in accuracy, circuit complexity, and trainability.
2 METHODOLOGY
- The pipeline uses Haar cascades to detect faces under controlled lighting and frontal or near-frontal capture, extracts rectangular ROIs, resizes them to 128×128, applies gamma correction and CLAHE; ROI extraction helps reduce inference time.CLAHE used a 2.0 clip limit and 8 × 8 tiles; gamma correction used γ = 2.2, selected after testing values from 1.8 to 2.6.Source: §2 · p. 5
- The authors report that CLAHE produced cleaner images with facial features visible to human observers and the quantum classifier.CLAHE limits contrast amplification while applying histogram equalization adaptively across image tiles.Source: §2 · pp. 5–6
- PCA reduces each flattened 16,384-dimensional image to eight components for quantum encoding.The eight-component choice is described as balancing classification accuracy and fewer quantum-encoded inputs.Source: §2 · p. 6
- The VQC uses eight qubits with angle encoding; the architecture also includes strongly entangling layers and measurement.Figure 2 describes angle embedding, circular nearest-neighbor CNOT entanglement, and Pauli-Z expectation-value readout.Source: §2 · p. 6, Figure 2 · p. 7
- Angle encoding performed about 1% better than amplitude encoding while reducing circuit complexity, and the authors report 87% accuracy for angle encoding.They also describe angle encoding as robust to noise and note that amplitude-state preparation requires deeper circuits with more gates.Source: §2 · p. 7
- Three ansatz layers were selected as an accuracy–convergence trade-off; adding layers reduced accuracy and trainability.The final circuit has 72 trainable parameters, compared with 2.3 million in the FaceNet baseline.Source: §2 · p. 7
3 Results and Discussion
Across the 10,000-image test set, the VQC outperformed the CPU FaceNet and QSVM baselines in accuracy and training time. The results also include real-quantum-hardware tests and attendance deployment, while leaving quantum advantage unresolved.
3.1 Dataset Construction
- The synthetic evaluation used 40,000 training images and 10,000 test images, balanced between faces and non-faces.The broader synthetic datasets were CelebA and CIFAR-10; real-world deployment used a dataset of 200 students.Source: §3.1 · p. 8
3.3 Classification Performance
- VQC achieved 97% accuracy, with 97% of non-faces and 95% of faces correctly classified; accuracy was 97.5% without spectacles and 96.8% with them.Inference took 0.2–0.5 seconds per image, and quantum-circuit execution accounted for 66.5% of total time.Source: §3.3 · pp. 9–11
3.4 Hardware Runtime Results
- On IBM Torino, increasing the shot count from 512 to 1024 reduced confidence-score standard deviation from 0.057 to 0.027, while the study reported stable decision behavior across repeated runs without retraining.These tests used the trained 8-qubit, 72-parameter VQC; increased shot counts also introduced execution overhead.Source: §3.4 · pp. 11–12
3.3 Classification Performance
- 97% accuracy and 0.91h training time for VQC compare with 85.1% accuracy and 243h for CPU FaceNet, and 80% accuracy and 38h for QSVM.All three approaches were tested on identical CPU-only hardware; Figure 5 summarizes the comparison.Source: §3.3 · p. 9, Figure 5 · p. 12
3.5 Interpreting of Results
- The authors state that quantum advantage remains unclear because highly optimized classical algorithms might match the reported performance.They suggest the 72-parameter circuit’s performance relative to a 2.3-million-parameter classical network as evidence of a possible quantum contribution.Source: §3.5 · p. 13
4 Limitations
The system is limited to binary face-versus-non-face classification, while identifying specific people requires extending it to N-class classification. The reported experiments use noiseless quantum simulation, so physical-hardware error effects remain unmeasured.
4 Limitations
- Real-world identification requires extending face-versus-non-face classification to N-class recognition, which the authors describe as challenging.They identify multiple binary classifiers or measuring more qubits as possible approaches, but note that both pose training challenges.Source: §4 · p. 14
- Multiple binary classifiers or measurements with more qubits are proposed for N-class recognition, but both approaches pose training challenges.The authors frame binary classification as a first step toward full identification.Source: §4 · p. 14
- Noiseless quantum simulation leaves performance degradation from physical-hardware errors unmeasured, making testing on actual quantum hardware future work.The authors cite decoherence, gate errors, and readout errors as potential sources of degradation.Source: §4 · p. 14
5 Conclusion
On a balanced 50,000-image task, the hybrid system outperformed the CPU-trained classical baseline in accuracy and training time, and it also achieved high recognition accuracy in attendance deployment. The authors conclude that such systems may be efficient alternatives for specific biometric tasks, while broader generalization and hardware validation remain open.
5 Conclusion
- 97% binary classification accuracy with training under one hour on standard CPU hardware exceeded the identical-hardware classical baseline’s 85.1% accuracy and longer training time.The evaluation used a balanced dataset of 50,000 images.Source: §5 · p. 14
- 96.2% recognition accuracy with 0.2–0.5 seconds inference latency per subject was reported in real-time attendance monitoring.This deployment result was presented as evidence of practical viability.Source: §5 · p. 14
- An eight-qubit architecture with three strongly entangling layers balanced representational power, gradient stability, and noise resilience.The ablations found that performance depended on qubit selection, encoding strategy, and ansatz depth.Source: §5 · p. 14
- Multi-class identification, physical quantum-hardware validation, and broader dataset diversity and scalability remain to be investigated.These are needed to assess hardware noise effects and generalization performance.Source: §5 · p. 14
- Carefully designed hybrid quantum–classical systems may offer computationally efficient alternatives for specific biometric tasks where high-performance GPUs are unavailable.The conclusion notes that advances in quantum hardware and error mitigation may further enhance practicality.Source: §5 · p. 15