EIGEN DECOMPOSITION Sample Clauses
EIGEN DECOMPOSITION. Eigen decomposition provides an attractive alternative to cross correlation matching. Although equivalent to normalised crosscorrelation, it provides speed and accuracy increases not available with the original correlation coefficient based technique. Eigen decomposition also enables matching to take place within a specialised matching space, based on the objects to be matched. This reduces the dimensionality of the matching process and effectively compresses the images. Originally, Eigen space matching focused on matching an entire object to another object. This required that the objects to be matched were complete and not occluded. Within the AVITRACK project this is quite unlikely, and were it to be assumed, many false matches would ensue. This makes the original technique unsuitable for use in this project, however, adaptations and variations upon the theme of Eigen matching, such as that discussed in [85] may be more applicable.
EIGEN DECOMPOSITION. It is possible to plot an image in high dimensional space, which for a standard 200*200 image, would have roughly 10.25 million possible points. This image plot could then be used to match other candidate images, and from this establish class ownership; however, this is a slow (due to the high dimensional nature of the image projection) and inaccurate system (due to the projection of noise, e.g. unwanted image sections). To minimise these problems, an Eigen space approach is taken. This approach finds the most important (principle) components of an object or set of objects and discards the noise. This system also reduces the dimensionality of the problem, thereby improving the accuracy and speed of the technique. In effect, the image is compressed, and then the compressed image is used for matching. In this case, and for the reasons given in Section 5.2, small sections (image subsections) may be used to classify the objects. This approach is known as an EigenWindow approach. To understand the eigen window approach, it is useful to understand the founding Eigenspace approach. This approach aims to reduce the dimensionality of high dimensional information using principle components analysis (PCA).