AS-DQCNN: Angle superposition-augmented decreasing quantum convolutional neural network for image classification

Driven by the rapid advances in quantum computing, various quantum convolutional neural networks (QCNNs) have been developed as the combination of quantum computing and classical convolutional neural network, which seek to exploit the properties of quantum computing, such as superposition and entanglement, to enhance feature extraction capabilities in image classification. However, in most of the existing QCNNs, the quantum pooling layer (QPL) halves the qubit count, giving rise to an exponential reduction of image features. We borrow a decreasing QCNN (DQCNN) binary classifier, in which the QPLs are constructed with decreasing parametric quantum circuits, decreasing the image features of each QPL by N2 instead of N. To effectively adapt to multi-class classification tasks with a potential challenge that an unbalanced dataset distribution can lead to the decrease of the model's classification performance, we design a robust DQCNN by introducing the angle superposition (AS) technology (termed as AS-DQCNN), which aims to generate a balanced dataset by superimposing all remaining classes as a synthetic negative class with an identified positive class. As a result, the AS-DQCNN model is suitable for both binary and multi-class classifications. Especially, the latter can be achieved by utilizing multiple DQCNNs (backbone) along with a one-versus-rest strategy given that the DQCNN backbone is defined as a quantum binary classifier. In addition, we verify the DQCNN backbone's performance by performing 45 binary classification tasks and evaluate AS-DQCNN's performance using multiple DQCNNs for multi-class classification tasks on MNIST and Fashion-MNIST datasets. Experimental results indicate the DQCNN backbone achieves strong accuracy and fast convergence in binary classification, while AS-DQCNN delivers superior and balanced multi-classification accuracy, with macro-average precision, recall, and F1-score all exceeding 95.70% on both datasets. Comparison experiments exhibit the DQCNN backbone demonstrates competitive performance in average accuracy and peak accuracy, and shows superior performance particularly on the more complex Fashion-MNIST dataset. Especially in comparison with the DQCNN backbone, the AS-DQCNN achieves Top-1 accuracy improvements of 3.45% on MNIST and 7.24% on Fashion-MNIST with comparable loss values, demonstrating the effectiveness of the AS method in addressing data imbalance, as confirmed by paired t-tests (p < 0.001 for both datasets). Finally, we assess the circuit complexity and noise sensitivity of AS-DQCNN, primarily implying its applicability in noisy intermediate-scale quantum devices with tolerable noise thresholds.