A regularized correlation search is a search method in which correlation coefficients (the degree of similarity) for a portion of an image and a template are calculated, and that similarity is used as a yardstick. By measuring the similarities, the feature position of an image with features that resemble the template can be identified, meaning a regularized correlation search can be used for position detection.
As a result of a regularized correlation search, X,Y coordinates and scores (correlation coefficients) are output. Generally, the repeat positional accuracy is one quarter of a pixel for 3σ.

The scores (correlation coefficients) are calculated according to the correlation functions below. Firstly, the image is defined as a discontinuous two-dimensional function I (x, y). The search image is I, the search image range is ON, the model image is M, the model image range is Oi, and the relationship between i and N is 0≦ i <N

Methods used to calculate the correlation coefficients are general mathematical calculations and are the same for each image processing manufacturer. However, if all the pixels and all the bit numbers are calculated, it takes a huge amount of processing time, so manufacturers use their own individual algorithms to reduce the calculations (= abbreviated processing). In addition, each company uses their own methods to select the most correct scores from among several candidates, so due to these two processing differences, processing speed and precision differ with each manufacturer.

Regularized correlation search applications

In many cases, two search places are registered, and slight angle deviations or the like are corrected.

Another example of an application for a regularized correlation search is to use correlation values to inspect for the mixing of different kinds of products and similar issues.