Image Analysis via Applied Harmonic Analysis : Perceptual Image Quality Assessment, Visual Servoing, and Feature Detection

Abstract

Certain systems of analyzing functions developed in the field of applied harmonic analysis are specifically designed to yield efficient representations of structures which are characteristic of common classes of two-dimensional signals, like images. In particular, functions in these systems are typically sensitive to features that define the geometry of a signal, like edges and curves in the case of images. These properties make them ideal candidates for a wide variety of tasks in image processing and image analysis. This thesis discusses three recently developed approaches to utilizing systems of wavelets, shearlets, and alpha-molecules in specific image analysis tasks. First, a perceptual image similarity measure is introduced that is solely based on the coefficients obtained from six discrete Haar wavelet filters but yields state of the art correlations with human opinion scores on large benchmark databases. The second application concerns visual servoing, which is a technique for controlling the motion of a robot by using feedback from a visual sensor. In particular, it will be investigated how the coefficients yielded by discrete wavelet and shearlet transforms can be used as the visual features that control the motion of a robot with six degrees of freedom. Finally, a novel framework for the detection and characterization of features such as edges, ridges, and blobs in two-dimensional images is presented and evaluated in extensive numerical experiments. Here, versatile and robust feature detectors are obtained by exploiting the special symmetry properties of directionally sensitive analyzing functions in systems created within the recently introduced alpha-molecule framework.