The rise of artificial intelligence (AI) as a general solution in image processing problems, such as segmentation and detection tasks, has sparked growing interest in generating synthetic image data. A significant factor is that real data is often not available in sufficient quantity or quality for the thorough development, evaluation, and analysis of solutions. The reasons for this are diverse: image acquisition can be costly, labeling may require unfeasible and error-prone manual annotation, or the data may have low variability that does not allow for generalization. With the increase in computational power over the last few decades, synthetic data has become an essential component in the development of solutions as it addresses the shortcomings of real data. Furthermore, it provides almost unlimited diversity, is easily acquired, and comes with perfect annotation. This thesis consists of three parts that highlight the significance of synthetic data in developing, validating, and analyzing methods and results in image processing. In the first part, we present a convex optimization method for removing stripe artifacts. These elongated and parallel corruptions appear frequently with various imaging techniques including light-sheet fluorescence microscopy (LSFM), focused ion beam scanning electron microscopy (FIB-SEM), and remote sensing. Our approach offers intuitive parametrization and is highly flexible to different scenarios of image structures and stripes. We demonstrate the effectiveness and advantages of our approach over existing solutions by comparing them across real images from LSFM, FIB-SEM, and remote sensing through visual inspection. Based on synthetic LSFM data obtained by simulating physical light propagation we enrich our analysis by comparing the processed images to ground truth data and quantitatively confirming the performance observed on real data. In the second part, we discuss the assessment of quality for results in binary image segmentation tasks by comparing a large variety of established traditional metrics and distance-based approaches. This includes a distance-based metric that we propose, which captures the spatial distribution of errors while offering desirable properties such as normalization and interpretability. Using predominantly synthetic data and some real segmentation results, we perform a thorough analysis of the segmentation metrics across diverse conditions. This demonstrates the robustness and effectiveness of our metric in distinguishing errors near the surface from those farther away across different structural contexts. We illustrate its inclusion to real-world segmentation tasks by extending a previous study on segmenting cracks in CT images. In the third part, we conduct a systematic study of morphologically diverse geometric structures with the goal to characterize the morphology of 3D spatial structures in terms of their anisotropy, scale and angularity based on the measurement of common morphological features. Geometries are entirely generated using stochastic models whose parameters intuitively translate to the studied concepts and yield morphologically diverse structures. Using classical machine learning approaches for dimensionality reduction our study finds a series of measures for anisotropy, scale and angularity based on linear combination of the morphological features that distinguish between synthetic structures accordingly. We exemplify and discuss their use on real data for modeling of general spatial structures.
Niklas Rottmayer (Thu,) studied this question.