Image registration is a fundamental procedure for analyzing data from multiple sources. In medicine, it combines anatomical images (MRI, CT) with functional ones (PET, EEG); in remote sensing, it aligns multispectral frames for environmental monitoring; in computer vision, it aligns frames for motion tracking. Alongside this work in the archive are articles on medical image segmentation, dimensionality reduction, and weakly supervised learning, all of which address related image analysis tasks but do not focus on geometric alignment. This review systematically organizes registration methods, filling a gap in the classification of approaches.
The author conducted a literature review of image registration methods. Four mandatory steps are identified: feature detection (closed regions, edges, corners), feature matching between images, estimation of transformation parameters, and image transformation. Methods are classified according to ten criteria: dimensionality, domain of transformation, type of transformation, quality, parameters, subject, object, nature of basis, interaction, and modalities. The review covers external markers, surfaces, moments and principal axes, correlation, mutual information, wavelets, and soft computing (neural networks, fuzzy sets, optimization heuristics). For accuracy assessment, the correlation coefficient and mutual information are proposed.
The review shows that there is no universal registration method: the choice depends on the type of images and distortions. For monomodal images, correlation works well; for multimodal images, mutual information is preferred, as it does not require the same nature of intensities. External markers provide high accuracy but are invasive; surface-based methods are suitable for rigid transformations; moment-based methods are fast but sensitive to segmentation. Soft computing, especially evolutionary algorithms, effectively searches for optimal parameters of affine transformations. Hybrid approaches combining mutual information with feature-based measures enhance robustness.
This work stands somewhat apart from neighboring articles that focus on segmentation and representation learning. Here, the focus is on geometric alignment, making the review a valuable reference for researchers who need to combine data from different sources. Unlike narrow methods, the review provides a holistic picture and highlights the open problem of automatic registration, especially under complex nonlinear distortions and occlusions.
Promising directions include the development of feature-based methods with modality-insensitive descriptors, which would ensure robustness and accuracy. Hybrid schemes combining mutual information with feature-based measures could improve reliability. It is also important to account for radiometric distortions and noise, not just geometric deformations.
Improving registration methods will directly impact medical diagnostics (fusion of MRI and PET), environmental monitoring via satellite imagery, and computer vision systems.
Develop new feature-based methods with modality-independent descriptors and test hybrid approaches on real multimodal data with noise and occlusions.
In image processing, the problem of automatic registration under complex nonlinear distortions, occlusions, and varying illumination remains unsolved. This review shows that existing methods are either specialized or require manual tuning, so creating a universal and robust algorithm is an open challenge.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The author reviewed the literature on image registration methods, identifying four stages: feature detection, feature matching, transform estimation, and transformation. The methods are classified according to ten criteria, including dimensionality, type of transformation, and modality. External markers, surfaces, moments, correlation, mutual information, wavelets, and soft computing are discussed.
Where its strength is
Systematizing registration methods across multiple criteria provides a comprehensive overview and highlights the absence of a universal solution, making this review a useful reference.
Where it could go next
Worth trying
Why this may matter
Image registration remains a key task when combining data from different sources, and advancing methods toward robustness against nonlinear distortions and occlusions could expand their application in medicine, cartography, and remote sensing.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.57, against a usual 0.64. It looks like few have drilled nearby.
For aligning MRI and PET brain images, mutual information is often used: it finds common patterns in intensities, even if one image shows anatomy and the other shows metabolic activity.