Eigen-css Shape Matching and Recognizing Fish in Underwater Video

This thesis presents work on shape matching and object recognition. First, we describe Eigen-CSS, a faster and more accurate approach to representing and matching the curvature scale space (CSS) features of shape silhouette contours. Phase-correlated marginal-sum features and PCA eigenspace decomposition via SVD differentiate our technique from earlier work. Next, we describe a deformable template object recognition method for classifying fish species in underwater video. The efficient combination of shape contexts with larger-scale spatial structure information allows acceptable estimation of point correspondences between template and test images despite missing or inaccurate edge information. Fast distance transforms and tree-structured dynamic programming allow the efficient computation of globally optimal correspondences, and multi-class support vector machines (SVMs) are used for classification. The two methods, Eigen-CSS shape matching and deformable template matching followed by texture-based recognition, are contrasted as complementary techniques that respectively suit the unique characteristics of two substantially different computer vision problems.