What Is Image Stitching?

Image stitching is a computational photography technique that combines two or more overlapping images into a single continuous composite, typically a wide-angle panorama or a higher-resolution mosaic. The process requires aligning the constituent images so that corresponding scene points coincide and then blending the overlapping regions so that transitions are not visible to the observer. Image stitching is also called photo mosaicking when the composited area extends over large spatial extents, as in satellite image mapping.

The discipline draws from projective geometry, feature detection, numerical optimization, and image blending theory. A camera rotating about its optical center captures views related by planar homographies, a property that underlies the validity of panoramic stitching. When camera geometry departs from this ideal, more general warping models such as cylindrical or spherical projections are applied. The core pipeline, regardless of the projection model, consists of three stages: feature matching, geometric registration, and photometric blending.

Feature Detection and Matching

Reliable stitching depends on identifying the same physical points across different images despite differences in scale, rotation, and illumination. Local feature detectors extract keypoints at visually distinctive image locations, and descriptor vectors encode the appearance neighborhood around each keypoint. The Scale-Invariant Feature Transform (SIFT) algorithm, introduced by David Lowe, produces descriptors that are invariant to scale and rotation and exhibit partial tolerance to affine distortion. SURF (Speeded-Up Robust Features, named for its computational efficiency) approximates SIFT with integral images for faster computation, while ORB (Oriented FAST and Rotated BRIEF) provides a binary descriptor suitable for real-time applications. Matching proceeds by comparing descriptor vectors from two images and retaining candidate matches; outliers introduced by repetitive textures or background clutter are filtered using RANSAC. The seminal paper on automatic panoramic stitching using invariant features by Brown and Lowe established the feature-based registration approach that most modern stitching systems follow.

Image Registration and Homography

Once correspondences are established, registration recovers the geometric transformation that maps one image onto the coordinate system of another. For planar scenes or cameras rotating about a fixed center, this transformation is a homography: a 3-by-3 matrix that encodes the full projective mapping between two views. Estimating the homography from matched keypoints uses the Direct Linear Transform (DLT) algorithm, typically embedded within RANSAC to discard erroneous matches. When images span a very wide field of view, cylindrical or spherical projection warps each image onto a common curved surface before combining, reducing parallax-induced misalignment. Bundle adjustment is applied in multi-image pipelines to simultaneously refine all camera parameters and reduce accumulated registration error across the full sequence. A comparative study of image stitching methods surveys how registration accuracy varies with scene type, camera model, and the number of images in the composite.

Blending and Compositing

After registration, stitched images exhibit intensity discontinuities at seam boundaries caused by exposure differences, vignetting, or slight misregistrations. Blending algorithms reduce these artifacts by creating smooth transitions in the overlap region. Feathering applies a linear intensity ramp across the overlap, weighted by distance from the seam. Multiband blending, developed in the context of panoramic compositing, decomposes images into a Laplacian pyramid and blends each frequency band separately, attenuating low-frequency color differences with broad transition zones while preserving sharp high-frequency details. Optimal seam finding selects the join line at pixels where the two images differ least, minimizing perceptible discontinuities. A unified framework for street-view panorama stitching describes how these components are combined at scale for mapping applications.

Applications

Image stitching has applications in a range of fields, including:

• Consumer photography: panoramic images captured by smartphone cameras and dedicated panoramic cameras
• Satellite and aerial mapping: assembling large-area ortho-mosaics from overlapping survey images for cartography and land analysis
• Virtual tours and real estate: constructing 360-degree interior views from sets of wide-angle photographs
• Medical imaging: stitching retinal fundus photographs to produce a wide-field map of the optic fundus
• Microscopy: assembling high-resolution tissue maps from tiled microscope fields of view