IMPROVING EIGEN WINDOW SELECTION. To assess the suitability of fragments (Eigen windows) used in the last section, metrics may be used. One such metric, based on mutual information, is proposed here. This metric, used by Xxxxxx, Xxxx and Xxxxx Xxxxxx [26], describes the amount of information a specific fragment or patch contains about the image in question. This measure is found by: I C , F =H C −H C / F (6.25) Where C represents the class as a whole and F represents the fragment. H(X) represents the entropy of X, and is found by using the standard entropy measure, as given in Equation 5.17. Once this metric is used, fragments (image patches) with high mutual information content are found. These are then ranked to resolve the patches with the best content. However, based on this metric alone it is not possible to extract patches which are likely to represent their respective class adequately. To locate patches which not only contain high mutual information concerning their class, but also represent their class with respect to classification, and therefore prove useful as a basis for class matching purposes, another metric must be used. This metric assesses the degree (based on probability) to which a certain patch is present within images of its own class as opposed to images of other class types. This is achieved through use of the likelihood ratio: x= p F∣C p F∣NC (6.26) where NC represents all classes other than the class the fragment originated from. Due to the length of time required for this type of search, pair wise testing may first be used to highlight patches, which were likely to be useful. These patches are then tested on a reduced set of class images to further focus the set of possible image patches, before a full analysis, which gives accurate class representation probability based upon the above likelihood ratio is used. Once the two measures are combined, patches with high informational content, and which represent their class, can be extracted and used as a basis for classification. During the matching phase of this technique (used when ascertaining the values of the likelihood ratio for each candidate patch), two types of matching scheme are used [26], namely normalised cross correlation and an ordinal ranking measure. Other measures, such as sum of squared differences (SSD), SSD=∑ I 1−I 2 2 , where I 1 , I 2 represent the intensity values in image patch 1 and 2, which is also often used, is not included here as it suffers from an increased sensitivi...
IMPROVING EIGEN WINDOW SELECTION. The information given in Section 6.4.3 concerning Eigen window selection was used to ensure the best possible Eigen windows were used in the classification phase. This was a slow and involved process, however, as the results show, improvement has been affected in the classification results due to this. Cross correlation is slow and inaccurate, and was therefore not used to assess the quality of patch sets. The use of the kmeasure provides a considerable speed increase over the cross correlation method; however, for extensive search problems, such as searching an entire object class of patches, the same efficiency dilemma reoccurs, as the images must be converted to ordinal rank representation.
