Scholarlink

<p>This study proposes a hybrid quantum self-attention network (HQSAN) to address the challenge of modeling long-range dependencies in remote sensing image classification, a task where conventional CNNs are often limited. The HQSAN integrates a quantum processing layer (QPL) with a quantum self-attention layer (QSA). The QPL employs amplitude encoding and angle encoding to project features into high-dimensional quantum state space, enabling global feature transformation through parameterized quantum circuits. The QSA leverages quantum entanglement and superposition effects, incorporating all-to-all, ring, and circuit block topological configurations combined with Hadamard gates and controlled rotation gates to achieve long-range dependency modeling. Simulation experiments on PatternNet, RSI-CB256, and RSI-CB128 datasets show promising classification accuracies of 98.09%, 98.57%, and 98.50% respectively, surpassing the performance of traditional models including VGG-16, ResNet, and EfficientNet in the same noiseless setting. This quantum–classical collaborative framework demonstrates the potential to enhance non-local feature interactions, which could significantly improve the accuracy of remote sensing image classification upon future hardware realization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hqsan: a hybrid quantum self-attention network for remote sensing image scene classification

  • 期刊论文

  • 发表日期：

    2025年10月8日

查看全文
  • Ruiqi Wang,
  • Zhongrong Zhang,
  • Yulin Shen,
  • Zhenghao Sui,
  • Kang Lin

摘要

This study proposes a hybrid quantum self-attention network (HQSAN) to address the challenge of modeling long-range dependencies in remote sensing image classification, a task where conventional CNNs are often limited. The HQSAN integrates a quantum processing layer (QPL) with a quantum self-attention layer (QSA). The QPL employs amplitude encoding and angle encoding to project features into high-dimensional quantum state space, enabling global feature transformation through parameterized quantum circuits. The QSA leverages quantum entanglement and superposition effects, incorporating all-to-all, ring, and circuit block topological configurations combined with Hadamard gates and controlled rotation gates to achieve long-range dependency modeling. Simulation experiments on PatternNet, RSI-CB256, and RSI-CB128 datasets show promising classification accuracies of 98.09%, 98.57%, and 98.50% respectively, surpassing the performance of traditional models including VGG-16, ResNet, and EfficientNet in the same noiseless setting. This quantum–classical collaborative framework demonstrates the potential to enhance non-local feature interactions, which could significantly improve the accuracy of remote sensing image classification upon future hardware realization.