The goal of the present work was to develop a system for automated similarity retrieval of figurative images---especially trademark images---which gives results that resemble human similarity estimation. 

In the first chapter, findings about the peculiarities of the perception of images and about human similarity estimation are compiled and the special needs of similarity retrieval of trademark images are explained.

As the depicted shapes play an important role for the estimation of similarity, an approach for the detection of the shapes has been developed. It encounters that shapes may be depicted in different ways (by regions, using textures, by contour lines) and that images often contain compression artefacts and noise.

For the estimation of the similarity of images based on the detected shapes, an approach has been developed that, in a first stage, computes transformations which map the images and, in a second stage, compares the mapped images. For the computation of the transformations an existing randomized approach has been enhanced. It chooses appropriate transformations based on collecting votes. For the comparison of the mapped images a new similarity measure on the contour lines has been developed which takes the correspondences in position and direction into account.

Based on these components a system for similarity retrieval has been developed which also considers the special needs of similarity retrieval of trademark images. The experimental results show a high conformance with human similarity estimation. The results are significantly better than the ones achieved by existing systems.