Low level image processing using steerable filters

The thesis is concerned with low level methods for the analysis of local image structures. The term 'local' refers to generic image structures that can be described data driven and mathematically. The term 'low level' refers to the first steps in a processing hierarchy from the gray value representation to the semantic interpretation of images. The first processing step thereby consists of the application of appropriate filters. This thesis especially investigates the signal processing properties of certain filters as well as a new filter technique. By using filter methods that are more complex than simple edge detectors, the range of recognized local image structures is significantly extended compared to 'standard' image processing systems. The thesis consists of the following three parts. The first part discusses the main principles the choice of the filters is based on. The Gabor functions and the derivatives of Gaussians are two classes of functions that are especially distinguished by these principles. One of these principles is the energy/phase representation that is obtained by quadrature filters. This principle is of special importance for the presented analyzing methods. Therefore, we investigate in detail the quality of the (pseudo-) quadrature filters that can be realized by Gabor functions or derivatives of Gaussians. In addition, we investigate the relationship between the two classes of functions. The second part presents the method of steerable filters. Steerable filters have been introduced recently to efficiently calculate the responses of filters in a continuum of scales, orientations, and other parameters. The extension and detailed investigation of different steering schemes is a major topic of this thesis. Beside the energy/phase representations, the efficient 'availability' of the filters for arbitrary orientations and scales is a second major input for the proposed image analysis methods. The third part is concerned with the analysis of local image structure. We steer a special double lobed pseudo quadrature filter in orientation for the analysis of junctions of edges and lines. From the number of significant maxima in the energy and the respective values of the phase, the number and type of edges and lines are determined. The double lobed filter shows good performance for analyzing junctions but the calculation of its responses is expensive. Motivated by this filter, we propose a two layer filter technique for a more efficient calculation of various complex filters. In a first layer a simple filter is steered in orientation and scale. The second layer synthesizes various complex filters by superpositions of the first layer filters. In an outlook we outline an attentive strategy based on the use of this filtering method. The attentive strategy, on the other hand, is a prerequisite for the use of complex filters. (orig./AFK)Available from TIB Hannover: RO 2674(1995,30) / FIZ - Fachinformationszzentrum Karlsruhe / TIB - Technische InformationsbibliothekSIGLEDEGerman