Feature extraction without edge detection
Ronald D. Chaney · DSpace@MIT (Massachusetts Institute of Technology) · 1993
Information representation is a critical issue in machine vision. The representation strategy at the primitive stages of a vision system has enormous implications for the processing capabilities in the subsequent stages. Existing feature extraction paradigms, like edge detection, provide sparse and unreliable representations of the image information. In this thesis, we propose a novel feature extraction paradigm. The paradigm is based on the dual interpretation of the Laplacian of Gaussian (LoG) as a matched filter and an edge locator. The zero-crossings of the LoG tend to outline subjective features in the image, such as isobrightness regions. The typical size of the outlined regions depends on the spatial width of the LoG filter. Hence, a naive approachwould be to take the regions bounded by the zero-crossings of the LoG filter as the features. In practice, such regions consist of multiple subjective regions that have merged together due to the smoothing process. To address this issue, weintroduce a stable, robust decomposition of regions into their salient parts. The resulting subregions, called simple region features, serve as the feature primitives for higher level processing. The region decomposition is computed from the medial axis skeleton of each region bounded by zero-crossings. Each subregion corresponds to a portion of a branch of the medial axis skeleton# each skeleton branch is divided at positions where the distance from the skeleton to the bounding contour is minimized. To facilitate the computation of the decomposition, a number of computational geometry problems are addressed. A novel scale-space is introduced for contours and the medial axis skeleton. The scale-space is parametric with the complexity of the contour or the skeleton. The complexity measur...