Continuous boundary reconstruction from noisy images using spatial interaction and application

Abstract

The thesis deals with methodologies for reconstructing meaningful boundary information from highly noisy images. Particular interest is put on defining families of edges which are faithful to actual 3D scene (good accuracy), but avoid non-physical ones caused by sensor noise, texture, unexpected lighting, etc. (good detection and robustness). To achieve the goals, a boundary reconstruction framework is formulated as a single optimization problem, which fuses the detection of edges with their pruning, linking and smoothing. Underlying the framework is iterative operation of adaptive spatial interaction between estimated neighboring image features, which selectively penalizes violation of prior knowledge on relationship between neighboring features to enhance both detection accuracy and detectability. Different three reconstruction models obeying this kind of framework are investigated in this thesis: (1) edge relaxation model, (2) robust line process model, and (3) elastic shape model. In edge relaxation model, candidates of edges first extracted by simple differentiation are refined in order to get more completely described boundaries by using statistics model for possible edge configurations in actual 3D scene. The statistics are obtained from actual input images through inductive learning by a specially designed neural network. Robust line process model is developed with probabilistic belief that most physical measurements except for those around edge points are reasonably smooth. Differentiation of smoothed image with discontinuities sharply preserved by robust line process guarantees accurate and reliable reconstruction of edge points. While this kind of two reconstruction models are totally data-driven process, elastic shape model combines both data-driven process and goal-driven process so that extraction of complete boundaries can be efficiently achieved in a single framework incorporating high-level knowledges, such as shape or its variability, into det...