ABSTRACT
The detection of elongated structures like lines or edges is an essential component in semantic image analysis. Classical approaches that rely on significant image gradients quickly reach their limits when the structure is context-dependent, amorphous, or not directly visible. This study introduces a principled mathematical description of elongated structures with various origins and shapes. Among others, it serves as an expressive operational description of target functions that can be well approximated by Convolutional Neural Networks. The nominal position of a curve and its positional uncertainty are encoded as a heatmap by convolving the curve distribution with a filter function. We propose a low-error approximation to the expensive numerical integration by evaluating a distance-dependent function, enabling a lightweight implementation with linear time complexity. We analyze the method's numerical approximation error and behavior for different curve types and signal-to-noise levels. Application to surgical 2D and 3D data, semantic boundary detection, skeletonization, and other related tasks demonstrate the method's versatility at low errors.
ABSTRACT
Oncologic therapy of laryngeal cancer may necessitate a total excision of the larynx which results in loss of voice. Voice rehabilitation can be achieved using mucosal tissue vibrations at the upper part of the esophagus which serves as substitute voice generating element (PE segment). The quality of the substitute voice is closely related to vibratory characteristics of the PE segment. By means of a high-speed camera the dynamics of the PE segment can be recorded in real-time. Using image processing the deformations of the PE segment are extracted from the image series as deforming contours. Commonly, the characterization of PE dynamics bases on the spectral analysis of the time varying contour area. However, this constitutes an integral approach which masks most of the specific dynamics of PE deformations. We present an algorithm that automatically registers one segmented contour in a frame of the video sequence to the contour in the next frame to derive discrete 2-D trajectories of PE vibrations. By concatenation of the obtained transformations this approach provides a total registration of PE segment contours. We suggest a mixed-integer programming formulation for the problem that combines an advanced outlier and deformation handling with the introduction of dummy points in regions that newly open up, and that includes normal information in the objective function to avoid unwanted deformations. Numerical experiments show that the implemented alternate convex search algorithm produces robust results which is demonstrated at the example of five high-speed recordings of laryngectomee subjects.