A Parameter-Efficient Hybrid Quantum-Classical Approach for Image Anomaly Classification

A promising paradigm for meeting the increasing computational demands of deep learning systems is quantum machine learning. Deep neural networks, which are highly parameterised and computationally demanding, are usually used for image-based anomaly classification, which is essential to applications in remote sensing, industrial inspection, and healthcare. To address these challenges, this work presents a parameter-efficient hybrid quantum–classical framework that integrates quantum computational principles with conventional deep learning architectures for efficient anomaly classification. The proposed approach combines a lightweight convolutional neural network (CNN) for feature extraction with an eight-qubit, two-layer Variational Quantum Circuit (VQC) acting as the classification module. The performance of the framework is tested on image-level anomaly classification and compared to the same classical CNN structure and two popular deep learning models, ResNet18 and MobileNetV2. The results of the comparison indicate that the hybrid CNN-VQC performs better in terms of Accuracy (84.29% vs. 87.18%), F1-score (78.22% vs. 82.30%), Precision (81.48% vs. 85.32%), Recall (75.21% vs. 79.49%), and AUC (88.93% vs. 91.81%) compared to its classical counterpart while decreasing the number of trainable classifier parameters from 257 to 32, which is a reduction of 87.5%. The hybrid CNN-VQC model maintains a stable learning process and the same level of prediction accuracy despite the drastic reduction in the number of trainable parameters. This experiment demonstrates the viability of hybrid quantum-classical models in creating lightweight and efficient anomaly classification systems that are suited to the constraints of near-term quantum computing hardware.