What Is Image Registration?

Image registration is the process of geometrically aligning two or more images of the same scene or subject so that corresponding points coincide. The aligned images may have been acquired at different times, from different viewpoints, by different sensors, or under different imaging conditions. Registration produces a spatial transformation that maps one image (the moving image) onto the coordinate system of another (the fixed or reference image), enabling direct comparison, fusion, or joint analysis of the data. The field draws on computational geometry, optimization theory, and pattern recognition, with applications spanning medical imaging, remote sensing, and computer vision.

A registration algorithm involves three main components: a deformation model that describes the family of allowable transformations, an objective function that measures how well the aligned images correspond, and an optimization method that searches for the transformation that maximizes correspondence. The choice of each component depends heavily on the imaging modality, the expected magnitude of misalignment, and whether the deformation is rigid (translation and rotation only) or non-rigid (local bending and stretching).

Feature-Based and Intensity-Based Methods

Registration methods are broadly divided into feature-based and intensity-based approaches. Feature-based methods first extract salient points, edges, or anatomical landmarks from both images and then compute the transformation that best aligns these features. SIFT (Scale-Invariant Feature Transform) and similar keypoint detectors are commonly used for multiview and satellite image registration, where distinctive texture elements provide reliable correspondence points. Intensity-based methods, by contrast, operate directly on the raw pixel or voxel values, using metrics such as cross-correlation, sum of squared differences, or mutual information to drive optimization. Mutual information is especially important in multimodal registration, where images from different sensors (for example, MRI and PET) have fundamentally different intensity distributions but carry complementary structural information.

Deformable Registration

Rigid alignment handles only global translations and rotations, which is sufficient for rigid structures such as bones. Many practical registration problems require deformable models that capture local shape variation, such as the motion of soft tissue during respiration or the anatomical differences between individuals in a population study. Deformable medical image registration, surveyed comprehensively in IEEE Transactions on Pattern Analysis and Machine Intelligence, covers the full taxonomy of deformation models: parametric models (B-splines, radial basis functions) that represent the displacement field with a compact set of control points, and non-parametric models (viscous fluid, diffusion) that allow nearly unconstrained local deformation. Deep learning has produced a new class of deformable registration networks, such as VoxelMorph, that learn to predict a dense displacement field directly from the input image pair in a single forward pass, reducing inference time from minutes to seconds compared to iterative optimization. Deep learning methods for medical image registration provide a comprehensive review of supervised, unsupervised, and weakly supervised architectures for this task.

Multimodal and Remote Sensing Registration

Multimodal registration aligns images acquired by instruments sensitive to different physical quantities, such as CT (X-ray attenuation), MRI (proton spin density and relaxation), and PET (metabolic tracer uptake). Because intensity patterns differ across modalities, similarity metrics based on information-theoretic measures are preferred. In remote sensing, registration aligns satellite or aerial images acquired at different times or by different instruments, enabling change detection, land-cover mapping, and the construction of orthorectified mosaics. Deformable registration with generative adversarial networks represents a line of work that has substantially improved accuracy for cross-modality alignment.

Applications

Image registration has applications in a wide range of fields, including:

• Longitudinal medical studies tracking tumor growth or treatment response
• Surgical planning through fusion of pre-operative MRI and intra-operative ultrasound
• Satellite image mosaicking and change detection in land-use monitoring
• Augmented reality and object tracking in robotics and autonomous systems
• Population brain atlases for neuroimaging research