Research map / Architectures and efficiency / CNN design
Content-Adaptive Non-Local Convolution for Remote Sensing Pansharpening
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CANConv, a content-adaptive non-local convolution that clusters pixels via K-Means and generates per-cluster kernels, and CANNet, a U-Net built from it for remote sensing pansharpening.
Problem
Pansharpening methods fail to exploit non-local self-similarity in remote sensing images, and spatial adaptive convolutions are redundant (one kernel per pixel) and local-only.
Approach
Cluster pixels globally by neighborhood features (similarity relationship partition), then generate one adaptive convolution kernel and bias per cluster from its centroid (partition-wise adaptive convolution), so all similar pixels anywhere in the image share a kernel.
Key results
- On WV3 reduced-resolution, CANNet achieves SAM 2.930, ERGAS 2.158, Q8 0.920, outperforming the next best (DCFNet/HMPNet) on most metrics.
- On WV3 full-resolution, CANNet reports HQNR 0.951 and Ds 0.0301, the best among compared methods.
- Plugging CANConv into DiCNN, FusionNet, and LAGConv improves SAM and ERGAS in each case (e.g., LAGConv SAM 3.104 to 3.051).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CANNet instead of the strongest listed baseline | CANNet · WV3 reduced-resolution | SAM: 3.038±0.585 -> 2.930±0.593 | Helped |
| use CANNet instead of the strongest listed baseline | CANNet · WV3 full-resolution | Dλ: 0.0165±0.0074 -> 0.0196±0.0083 | Hurt |
| use CANNet instead of the strongest listed baseline | CANNet · QB reduced-resolution | ERGAS: 3.404±0.478 -> 3.652±0.327 | Hurt |
| use CANNet instead of the strongest listed baseline | CANNet · GF2 reduced-resolution | SAM: 0.786±0.148 -> 0.707±0.148 | Helped |
| replace some backbone convolutions with CANConv | DiCNN · WV3 reduced-resolution | SAM: 3.593±0.762 -> 3.371±0.694 | Helped |
| replace some ResBlocks with CAN-ResBlocks | FusionNet · WV3 reduced-resolution | SAM: 3.325±0.698 -> 3.293±0.703 | Helped |
| replace some spatial-adaptive ResBlocks with CAN-ResBlocks | LAGConv · WV3 reduced-resolution | SAM: 3.104±0.559 -> 3.051±0.617 | Helped |
| disable SRP and use K = 1 (ablation) | CANNet · WV3 reduced-resolution | SAM: 2.930±0.593 -> 3.023±0.606 | Hurt |
| disable SRP and use K = ∞ (ablation) | CANNet · WV3 reduced-resolution | SAM: 2.930±0.593 -> 2.982±0.608 | Hurt |
| replace fk with a direct MLP (ablation) | CANNet · WV3 full-resolution | convergence | Hurt |
| set η = 0 when training (ablation) | CANNet · WV3 full-resolution | Ds: 0.0301±0.0074 -> 0.0400±0.0053 | Hurt |
Surprising findings
- Replacing the kernel generator with a plain MLP of the same depth increased learnable parameters about 10x and the network failed to converge.
Limitations
- K-Means adds computational cost (O(tKnd)) on top of adaptive convolution.
- Optimal cluster number K differs between training and test sets (about four times larger at test), requiring tuning.
Experimental setup
Datasets: WorldView-3 (WV3) pansharpening data via PanCollection QuickBird (QB) pansharpening data via PanCollection GaoFen-2 (GF2) pansharpening data via PanCollection
Benchmarks: WV3 reduced-resolution WV3 full-resolution QB reduced-resolution GF2 reduced-resolution
Metrics: SAM ERGAS Q8 Q4 Dλ Ds HQNR
Baselines: EXP MTF-GLP-FS TV BDSD-PC CVPR2019 LRTCFPan PNN PanNet DiCNN FusionNet DCFNet MMNet
Builds on: DFN PAC LAGConv CPNet IGNN
Closest papers by meaning
- Adaptive Rectangular Convolution for Remote Sensing Pansharpening (CVPR 2025 · similarity 0.89)
- A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · similarity 0.88)
- SSDiff: Spatial-spectral Integrated Diffusion Model for Remote Sensing Pansharpening (NeurIPS 2024 · similarity 0.87)
- Dual-Granularity Semantic Guided Sparse Routing Diffusion Model for General Pansharpening (CVPR 2025 · similarity 0.87)
- Pan-LUT: Efficient Pan-sharpening via Learnable Look-Up Tables (NeurIPS 2025 · similarity 0.86)
- HetSSNet: Spatial-Spectral Heterogeneous Graph Learning Network for Panchromatic and Multispectral Images Fusion (ICML 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.