Superposition-enhanced quantum neural network for multi-class image classification

Quantum neural networks have made progress in classification tasks. However, they face challenges when applied to multi-class image classification tasks. In this paper, we propose a superposition-enhanced quantum neural network(SEQNN). Comprising image superposition and quantum binary classifiers(QBCs), SEQNN addresses the following challenges. Firstly, the inherent linearity of quantum evolution is overcome by the one-vs-all strategy combined with QBCs, thereby circumventing the nonlinearity. Subsequently, the second challenge pertains to data imbalance within the subtasks of the one-vs-all strategy. Drawing inspiration from the mixup technique, image superposition is employed to alleviate this imbalance. Two image superposition methods, quantum state superposition(QSS) and angle superposition(AS), are proposed. The simulated experiments on MNIST and Fashion-Mnist show that AS is better than QSS in multi-class image classification tasks. Equipped with AS, SEQNN outperforms existing models and achieves an accuracy of 87.56% on MNIST.