Hierarchical gradient modulation for multi-resolution image registration

Abstract

In image registration, traditional methods often require manual supervision or the use of paired image data to optimize transformations, which can be both labor-intensive and limited in generalization. Unsupervised methods, by contrast, aim to automatically learn the optimal alignment between images without relying on labeled data, but they often struggle with balancing the complex trade-off between regularization and similarity loss, leading to issues with model tuning and weak generalization across diverse datasets. In this paper, a novel hierarchical gradient modulation strategy is proposed for multi-resolution image registration. This method introduces a compatibility criterion based on the relationship between the gradients of similarity loss and regularization loss. It evaluates the compatibility between similarity and regularization gradients, integrates them through intuitive strategies that align mutually reinforcing gradients, project conflicting gradients orthogonally to avoid interference, and balance gradients of equal importance through averaging. It prioritizes global deformation with stronger regularization at low resolution and focuses on fine details with reduced regularization at high resolution. Experimental results on common medical datasets, forward-looking sonar datasets, and fabric defect detection datasets demonstrate that the proposed method achieves superior registration performance compared to baseline methods and state-of-the-art hyperparameter-related research without incurring additional computational costs, achieving optimal loss balance in a multi-resolution structure.