Bildkorrrelation

Bildkorrrelation, often referred to in English as image correlation, denotes a family of techniques used to measure similarity between images or image regions. It is employed to determine relative displacement, orientation, rotation, or scale between images, and is fundamental to tasks such as image registration, object tracking, and template matching.

The core idea is to compute a correlation function between two images or patches. The basic cross-correlation at a shift (u, v) is C(u, v) = sum over x,y of I1(x, y) · I2(x+u, y+v). Variants such as normalized cross-correlation (NCC) normalize by means and variances to reduce effects of lighting changes. Phase correlation uses Fourier transforms to estimate translation from the peak of the inverse Fourier transform of the normalized cross-power spectrum, and is particularly robust to uniform illumination differences and noise.

Other common measures include the sum of squared differences (SSD) and the sum of absolute differences (SAD), which are often used in template matching. Multi-scale, hierarchical, or sub-pixel refinement approaches improve accuracy for small displacements or noisy data.

Applications span multiple fields. In computer vision, image registration aligns images from different times or sensors. In stereo vision, correlation methods aid depth estimation. In video analysis, they support motion tracking. Digital image correlation (DIC) extends these ideas to measure deformation and strain in materials by tracking speckle patterns across images.

Limitations include sensitivity to occlusion, rotation, or scale changes, as well as severe lighting variation and noise. Computational cost can be high for large images, though fast Fourier transform implementations and multi-resolution strategies mitigate this. In German literature, the term Bildkorrelation is used, sometimes alongside Digital Image Correlation (Digitale Bildkorrelation) for materiaI strain measurements.