GLAM: Global-Local Affine-Flow Matcher

• The paper introduces GLAM, a hybrid dense image registration module that combines global affine registration with fine-grained residual flow estimation for pixel-level SAR-optical alignment.
• It utilizes a coarse-to-fine cascade leveraging ResNet50-based feature pyramids and multi-scale gradient enhancement to capture shared geometric structures across modalities.
• Empirical evaluations show a +58.68 percentage point uplift in 1-pixel correspondence accuracy on the SEN1-2 dataset, validating GLAM's robust performance.

The Global-Local Affine-Flow Matcher (GLAM) is a hybrid dense image registration module designed for precise pixel-level alignment between synthetic aperture radar (SAR) and optical images. Developed within the SOMA framework, GLAM addresses the two critical alignment challenges intrinsic to cross-modal registration: maintaining structural global consistency while permitting fine-grained local corrections. It serves as the core refinement engine that, in conjunction with structurally enhanced deep features, aligns cross-domain feature pyramids via a coarse-to-fine architecture utilizing both affine deformation and residual flow fields (Wang et al., 17 Nov 2025).

1. Motivation and Underlying Principles

Traditional image registration techniques are generally insufficient for SAR-optical matching due to drastically different imaging modalities—SAR being sensitive to surface geometry and roughness, while optical imagery is radiometrically consistent and visually interpretable. Both, however, share underlying geometric structures such as edges, corners, and ridges, which can be robustly extracted and matched. GLAM is formulated to exploit these shared structures by first performing coarse global alignment using affine transformations and then refining registration through dense flow fields. This two-stage approach leverages the strengths of both global parametric models (for handling scale, rotation, and skew) and local non-parametric models (for non-rigid misalignments due to terrain, sensor distortions, or temporal change).

2. Architectural Overview and Operational Workflow

Within SOMA, both SAR and optical images are processed by separate ResNet50-based feature pyramid extractors, resulting in multi-resolution feature sets $\{\varphi_l\}$. GLAM is applied sequentially across scales in a coarse-to-fine cascade. At each resolution level $l$, the following operations are performed:

• The SAR feature map is warped using the deformation field estimated at the previous (coarser) scale.
• An affine registration module regresses a global affine field $\hat{W}^a_l$ representing scale, rotation, translation, and shear across the current pyramid level.
• A dense flow estimation module regresses a residual flow field $\hat{W}^f_l$ capturing local misalignments.
• The two transformations are composed to yield the current level’s deformation field; this field warps the SAR feature for subsequent finer-level matching.

This architecture ensures that coarse geometric misalignments are corrected globally, while local deformations are resolved by spatially-varying flow, facilitating high-precision registration at the finest level.

3. Mathematical Formulation

Let $F^{SAR}_l, F^{OPT}_l$ denote the structurally enhanced features (via Feature Gradient Enhancer, FGE) from SAR and optical images at pyramid level $l$, and let $\hat{W}_{l+1}$ be the deformation field propagated from the previous scale. GLAM's core computations are:

1. Warping: The SAR feature is spatially transformed: $$F^{SAR}_l \leftarrow \mathrm{Warp}(F^{SAR}_l, \hat{W}_{l+1})$$.
2. Affine Regression: Estimate a locally adaptive affine transformation field $\hat{W}^a_l$ using a regression network over $F^{SAR}_l$ and $F^{OPT}_l$.
3. Flow Regression: Using concatenated or difference features, a convolutional network predicts the residual flow field $\hat{W}^f_l$.
4. Field Composition: The final deformation is $\hat{W}_l = \hat{W}^a_l \circ \hat{W}^f_l$ (composition operator represents field addition in pixel space).

At the finest level $l=1$, the final dense deformation field $\hat{W}_1$ yields the pixelwise registration map from optical to SAR coordinates.

4. Multi-Scale and Coarse-to-Fine Strategy

GLAM’s cascade is critical for avoiding local minima typical in high-dimensional, non-rigid registration. The procedure is initiated at the coarsest scale:

• At $l=16$, a robust global feature anchor (e.g., from frozen DINOv2) facilitates overall registration initialization.
• Intermediate levels ($l=8,4,2$) receive successively refined deformation inputs, each time executing global affine correction followed by flow-based refinement.
• This strategy progressively restricts search space at each finer level, concentrating on correcting high-frequency misalignments not resolved at coarser resolutions.

5. Integration with Structural Priors and Feature Enhancement

GLAM’s effectiveness is amplified by preceding feature enhancement mechanisms within SOMA. The Feature Gradient Enhancer module embeds multi-directional, multi-scale Sobel-like gradient responses into intermediate deep features, selectively fused through channel and spatial attention. This process yields structurally distinctive and modality-robust representations ($\varphi_l^g$) ideal for dense correspondence estimation. As a result, GLAM operates not on generic CNN activations but on feature maps explicitly enriched for shared geometric structures across SAR and optical domains (Wang et al., 17 Nov 2025).

6. Empirical Impact and Practical Considerations

GLAM, within the SOMA framework, produces marked improvements in registration accuracy on benchmark datasets. For example, adding GLAM (combined with FGE and DINOv2) yields a +58.68 percentage point uplift at the 1-pixel correspondence accuracy (CMR@1px) on the SEN1-2 dataset compared to baseline approaches. The average root mean squared error is similarly reduced. GLAM incurs modest computational overhead and benefits from end-to-end differentiable design—gradients from matching losses backpropagate through both affine and flow branches, ensuring all components (including the feature enhancer) are optimized jointly.

7. Limitations and Context Within the Field

GLAM assumes that structural gradients are extractable and meaningful at intermediate feature resolutions. In scenarios of extremely weak texture or intense speckle, the affine plus flow paradigm may suffer, although multi-stage denoising and adaptive attention mitigate such risks. A plausible implication is that while GLAM is effective on remote sensing datasets with salient manmade or natural structures, generalization to textureless or highly dynamic scenes may require further architectural adaptation or domain-specific priors. Within the broader context, GLAM demonstrates how hierarchical registration strategies—previously common in traditional multi-scale optimization—are being re-integrated into modern deep learning pipelines for cross-modal correspondence (Wang et al., 17 Nov 2025).