HQF-Net: A Hybrid Quantum-Classical Multi-Scale Fusion Network for Remote Sensing Image Segmentation
In the realm of remote sensing, effective semantic segmentation is pivotal for interpreting complex scenes. Traditional models, particularly classical encoder-decoder architectures like U-Net, have demonstrated robust performance; however, they often fall short in leveraging global semantics and structured feature interactions. Addressing these limitations, a new model named HQF-Net has been introduced, combining quantum and classical methodologies to enhance remote sensing image segmentation.
Overview of HQF-Net
HQF-Net stands as a hybrid quantum-classical multi-scale fusion network designed specifically for remote sensing image segmentation. This innovative framework integrates advanced multi-scale semantic guidance through a frozen DINOv3 ViT-L/16 backbone alongside a tailored U-Net architecture. A key feature of HQF-Net is the Deformable Multiscale Cross-Attention Fusion (DMCAF) module, which adeptly merges information at various scales to improve segmentation accuracy.
Key Features and Innovations
The architecture of HQF-Net is characterized by several cutting-edge components:
- Deformable Multiscale Cross-Attention Fusion (DMCAF): This module enhances feature interaction by allowing the model to focus on relevant spatial regions across multiple scales.
- Quantum-Enhanced Skip Connections (QSkip): These connections facilitate improved feature refinement by integrating quantum principles, allowing for better information flow throughout the network.
- Quantum Bottleneck with Mixture-of-Experts (QMoE): This mechanism employs complementary local, global, and directional quantum circuits within an adaptive routing framework, optimizing the processing of features.
Performance Evaluation
To assess the effectiveness of HQF-Net, extensive experiments were conducted across three prominent remote sensing benchmarks. The results demonstrated significant improvements relative to traditional methods:
- On the LandCover.ai dataset, HQF-Net achieved a mean Intersection over Union (mIoU) of 0.8568, along with an overall accuracy of 96.87%.
- For the OpenEarthMap dataset, the model secured a mIoU of 71.82%.
- In the case of the SeasoNet dataset, HQF-Net recorded a mIoU of 55.28% and an impressive overall accuracy of 99.37%.
Additionally, an architectural ablation study was performed to validate the contribution of each major component of HQF-Net. This analysis confirmed that the structured hybrid quantum-classical processing significantly enhances the model’s performance in remote sensing semantic segmentation.
Conclusion
The introduction of HQF-Net marks a promising advancement in the field of remote sensing. By harnessing the strengths of both quantum and classical methodologies, this model paves the way for improved segmentation capabilities. As research continues, the integration of structured hybrid quantum-classical feature processing is anticipated to unlock new potentials for remote sensing applications, particularly under the constraints of near-term quantum technology.