(Georgia Institute of Technology, 2000-06)
Howard, Ayanna M.; Padgett, Curtis; Brown, Kenneth R.
This paper presents an algorithm for detecting a specified set of target objects embedded in visual imagery for an Automatic Target Recognition (ATR) application. ATR involves processing images for detecting, classifying, and tracking targets embedded in a background scene. We address the problem of discriminating between targets and non-target objects located within a cluttered environment by evaluating 40x40 image blocks belonging to a segmented image scene. Using directed principal component analysis, the data dimensionality of an image block is first reduced and then clustered into one of n classes based on a minimum distance to a set of n cluster prototypes. Following clustering, each image pattern is fed into an associated trained neural network for classification. A detailed description of our algorithm will be given in this paper. Evaluation of the overall algorithm demonstrates that our detection rates approach 96% with a false positive rate of less than 0.03%.
(Georgia Institute of Technology, 1999-03)
Howard, Ayanna M.; Padgett, Curtis; Brown, Kenneth R.
Many applications that use hyperspectral imagery focus on detection and recognition of targets that occupy a portion of a hyperspectral pixel. We address the problem of sub-pixel target detection by evaluating individual pixels belonging to a hyperspectral image scene. We begin by clustering each pixel into one of n classes based on the minimum distance to a set of n cluster prototypes. These cluster prototypes have previously been identified using a modified clustering algorithm based on prior sensed data. Associated with each cluster is a set of linear filters specifically designed to separate signatures derived from a target embedded in a background pixel from other typical signatures belonging to that cluster. The filters are found using directed principal component analysis which maximally separates the two groups. Each pixel is projected on this set of filters and the result is fed into a trained neural network for classification. A detailed description of our algorithm will be given in this paper. We outline our methodology for generating training and testing data, describe our modified clustering algorithm, explain how the linear filters are designed, and provide details on the neural network classifier. Evaluation of the overall algorithm demonstrates that for pixels with embedded targets taking up no more than 10% of the area, our detection rates approach 99.9% with a false positive rate of less than 10 ⁻⁴.