Hybrid Quantum-Classical Neural Networks for Analyzing Magnetic Tile Images for Defects

Research in the field of quantum computing encompasses the search for practical applications and the development of algorithms capable of delivering tangible benefits when executed on quantum computers of the noisy intermediate-scale quantum (NISQ) era. One promising area of application is the use of quantum, hybrid quantum-classical, or quantum-inspired algorithms for image classification tasks involving data of various origins. In this study, we address the problem of defect analysis in images of magnetic tiles through binary and multiclass classification problems. The proposed algorithms are based on quantum machine learning methods employing hybrid quantum-classical neural networks. As a baseline, we use classical convolutional neural network (CNN) in which one of the layers (specifically, the penultimate one) is replaced with a quantum layer. Experimental results obtained using both an ideal quantum simulator and a real quantum processor (ibmₛherbrooke) simulator demonstrate that the proposed solution achieves accuracy comparable to or exceeding that of a purely classical CNN, while requiring fewer trainable parameters. Moreover, the hybrid quantum-classical models exhibit greater robustness to overfitting, and the combination of quantum and classical layers slightly improves classification accuracy. Given the high number of quantum circuit runs required for training on real quantum systems, we additionally employ a two-stage training strategy: initial training is performed on an ideal quantum computer simulator, followed by fine-tuning on a real quantum computer simulator. In the context of our task, pre-training required several dozen epochs, while fine-tuning involved only a few epochs. Overall, these findings support the feasibility of applying quantum computers to practical classification problems, despite their current hardware limitations.