Julie Ann Gordon Whitney

Wavelet-based figure of merit for macro-uniformity

Wavelets are a powerful tool that can be applied to problems in image processing and analysis. They provide a multi-scale decomposition of an original image into average terms and detail terms that capture the characteristics of the image at different scales. In this project, we develop a figure of merit for macro-uniformity that is based on wavelets. We use the Haar basis to decompose the image of the scanned page into eleven levels. Starting from the lowest frequency level, we group the eleven levels into three non-overlapping separate frequency bands, each containing three levels. Each frequency band image consists of the superposition of the detail images within that band. We next compute 1-D horizontal and vertical projections for each frequency band image. For each frequency band image projection, we develop a structural approximation that summarizes the essential visual characteristics of that projection. For the coarsest band comprising levels 9,10,11, we use a generalized square-wave approximation. For the next coarsest band comprising levels 6,7,8, we use a piecewise linear spline approximation. For the finest bands comprising levels 3,4,5, we use a spectral decomposition. For each 1-D approximation signal, we define an appropriate set of scalar-valued features. These features are used to design two predictors one based on linear regression and the other based on the support vector machine, which are trained with data from our image quality ruler experiments with human subjects.

Expert-prescribed weighting for support vector machine classification

Support Vector Machines (SVM) are a family of algorithms that are used in classification and regression tasks. Often, multiple SVMs are combined in a coding scheme to provide multi-class classification capabilities. Generally, multi-class classification systems are evaluated on their accuracy of producing a correct coding by using test data and successful predictions are counted as a percentage of the whole, assuming that the test data set is a “good” representation of what the classification algorithm will see in its applied use. However, in practical applications, there may be situations where certain mistakes/confusions in classification are inconsequential to system operation. In this work, a method for integration of expert-defined allowable confusions into SVM systems is introduced, with an example implementation in a least squares support vector machine (LS-SVM) tested on industrial data, and shown to improve overall performance of a multi-class classification system when an appropriate performance measurement method is formulated.