Developing a validation metric using image classification techniques
Murali Mohan Kolluri · OhioLink ETD Center (Ohio Library and Information Network) · 2014
The main objective of this thesis work was to investigate different image classification and pattern recognition methods to try to develop a validation metric. A validation metric is a means of comparison between two sets of numerical information. The numerical information could represent a set of measurements made on a system or its internal characteristics derived from such measurements. A validation metric (v-metric) is used to determine the correctness with which one of the data-sets is able to describe the other and to quantify the extent of this correctness. A moment descriptor method has been identified from among the most widely used image classification and pattern recognition methods as the system most likely to give way to an effective validation metric for reasons discussed in subsequent chapters. Different sets of Orthogonal Polynomials have been investigated as kernel functions for the aforementioned method to generate descriptors that depict the most significant features of the data-sets being compared. The algorithms developed as such have been verified using standard gray-scale and color images to establish their ability to reconstruct the image intensity function using a subset of the features extracted. The above Orthogonal Polynomials have then been used to extract features from two measured data-sets and means to develop a v-metric from these descriptors have been explored. A study of algorithms thus developed using different Orthogonal Polynomials has been made to compare their effectiveness as well as shortcomings as kernel functions for developing a v-metric. An alternate form of the existing two dimensional moments has been proposed to generate features that are more conveniently compared against each other. This method has been examined to determine its efficiency in reducing the amount of information that needs to be used in the final comparison for multiple pairs of data-sets. A way to effect such a comparison using singular values of the computed moment matrix has been proposed.