arXiv:2609.31915
Facial classification Using Hybrid Quantum Machine Learning
Roshan Babu Bandlapalli, Srinivas V Katakam, Jitendra Chougala, Ravi Kumar Kappagantu, Jayasri Dontabhaktuni
Abstract
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.
AI Summary
The paper investigates whether a hybrid quantum–classical pipeline can support facial biometrics on CPU-constrained hardware. Its learned task is binary face/non-face classification; identity recognition is handled separately by classical database matching. Haar-cascade detection, gamma correction, contrast enhancement, and PCA reduce images to eight features, which are angle-encoded into an eight-qubit variational classifier with three entangling layers and 72 trainable parameters. Experiments use 25,000 CelebA faces and 25,000 CIFAR-10 non-face images, split 80:20 for training and testing. Noiseless CPU simulation reportedly achieves 97.0% accuracy with 0.91 hours of training, versus 85.1% and 243 hours for a simplified, CPU-trained FaceNet-style baseline, and 80.0% and 38 hours for a quantum support-vector machine. The reported face and non-face recalls, 95% and 97%, imply 96% accuracy on the balanced test set rather than the stated 97%. Attendance deployment reportedly achieves 96.2% recognition accuracy with 0.2–0.5-second latency per person. Hardware evaluation on IBM Torino repeats inference for one facial region ten times at each shot budget. Increasing shots from 512 to 1,024 reduces confidence-score standard deviation from 0.057 to 0.027; dataset-level hardware accuracy is not reported. Stated limitations include the gap between binary classification and multiclass identification, predominantly noiseless-simulation evidence, hardware-noise sensitivity, and unresolved scalability and generalization. Performance drops to 60–70% under direct illumination and approximately 60% for large profile angles. The paper acknowledges that optimized classical methods might match its performance and that quantum advantage is not established.
Central claim may not hold: The main experiment establishes only face/non-face classification, while identity matching is unspecified and the claimed 96.2% deployment recognition accuracy lacks supporting experimental details. On the balanced 10,000-image test set, the reported class recalls imply 96.0% accuracy and precisions inconsistent with the classification report, rather than the advertised 97.0% accuracy. The claimed accuracy and training-efficiency advantage rests on an inadequately documented comparison between a simplified FaceNet baseline described as using 1,000 images and a VQC described as training on 40,000 images, without matched lightweight classical controls. The abstract's GPU-evaluation claim has no documented experimental results, and the quantum-hardware test measures repeated confidence scores for only one facial ROI rather than classification performance across a dataset. The paper reports completed IBM Torino experiments but elsewhere states that physical-hardware validation remains future work.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.