Enterprise AI Analysis

This section reviews existing segmentation models, particularly U-Net and its variants, noting their strong baseline performance but limitations in complex remote sensing. It also covers the evolution of Quantum Machine Learning (QML) in vision tasks, from Quantum Convolutional Neural Networks (QCNN) to hybrid quantum-classical designs. The section emphasizes the strategy of placing quantum circuits at bottlenecks for feature refinement under NISQ constraints and introduces Mixture-of-Experts (MoE) models as inspiration for adaptive computation.

HQF-Net is a hybrid U-Net based architecture integrating quantum circuits within skip connections and the bottleneck. It leverages a pre-trained DINOv3 Vision Transformer to guide a classical encoder for multi-scale feature learning. Key components include: Deformable Multiscale Cross-Attention Fusion (DMCAF) for aligning DINOv3 semantics with U-Net spatial features; Quantum-augmented Skip connections (QSkip) for enhanced feature transfer; and a Quantum bottleneck with Mixture-of-Experts (QMoE), which combines local, global, and diagonal quantum circuits via an adaptive routing mechanism to capture complementary feature dependencies. This section details the classical encoder, DMCAF, and the quantum modules.

Experiments on LandCover.ai, OpenEarthMap, and SeasoNet datasets demonstrate HQF-Net's superior performance in remote sensing semantic segmentation. The model consistently outperforms classical and hybrid quantum baselines, achieving state-of-the-art mIoU and overall accuracy. Qualitative analysis shows cleaner object boundaries and better preservation of small structures. An ablation study validates the incremental contribution of each component, with DMCAF providing the largest gain, and QSkip and QMoE further refining performance, confirming the effectiveness of the hybrid quantum-classical design.

HQF-Net, a novel hybrid quantum-classical multi-scale fusion network, achieves strong and consistent semantic segmentation performance across diverse remote sensing benchmarks. Its architecture, combining DINOv3-guided fusion, DMCAF, quantum-enhanced skip refinement, and a QMoE bottleneck, effectively captures multi-scale semantic context and fine spatial details. The results validate the promising direction of structured hybrid quantum-classical feature processing for remote sensing under near-term quantum constraints, with future work planned to explore larger benchmarks and quantum-based self-supervised learning.