Hybrid Quantum Neural Networks with Group Averaging for Image Classification

The growing real-world deployment of computer vision technologies imposes increasingly stringent requirements on the generalization and few-shot adaptation performance of image classification models. While quantum machine learning (QML) has shown considerable promise for image classification tasks, existing hybrid quantum neural networks (HQNNs) consistently exhibit limited generalization ability and weak few-shot learning performance. We investigate the integration of standard W-group averaging into two existing HQNN architectures, HQNN-Parallel and HQNN-Quanv. Each complete input image is transformed by the four elements of the W-group and processed through a shared-parameter feature extractor. The resulting quantum feature vectors or feature maps are averaged before the final classical classification layer. This construction introduces no additional trainable parameters, but requires four transformed-input evaluations per image. The resulting models are denoted as W-HQNN-Parallel and W-HQNN-Quanv, respectively. Systematic experiments on three classic image classification datasets show that the W-HQNN-Parallel model achieves an accuracy of approximately 91.56% on the full FashionMNIST dataset and 73.96% on the full CIFAR-10 dataset respectively, outperforming the original model by around 0.50 and 2.88 percentage points, while also yielding lower test loss than HQNN-Parallel on both datasets. In single-run reduced-data experiments, W-HQNN-Quanv achieved accuracies of 68.83% on MNIST and 64.19% on FashionMNIST, exceeding the corresponding baselines by 9.60 and 6.16 percentage points, respectively. These results provide empirical evidence of the effects of W-group averaging in the two investigated HQNN architectures under the evaluated experimental settings.