Recovering the Orientation of Textured Surfaces in Natural Scenes.
ILLINOIS UNIV AT URBANA COORDINATED SCIENCE LAB
Pagination or Media Count:
A perspective view of a slanted textured surface shows systematic changes in the density, area and aspect-ratio of texture elements. These apparent changes in texture element properties can be analyzed to recover information about the physical layout of the scene. However, in practice it is difficult to identify texture elements, especially in images where the texture elements are partially occluded or are themselves textured at a finer scale. To solve this problem, it is necessary to integrate the extraction of texture elements with the recognition of scene layout. This thesis presents a method for recovering the orientation of textured surfaces while simultaneously identifying texture elements. Candidate texture elements are constructed from overlapping circular regions of relatively uniform gray-level. The uniform circular regions are found by convolving the image with Laplacian-of-Gaussian masks over a range of scales, and comparing the convolution output to that expected for a circular disk of constant gray level. True texture elements are selected from the set of candidate texture elements by finding the planar surface that best predicts the properties of the candidate texture elements. Each planar fit is evaluated by comparing the predicted texture-element areas to the actual areas of the candidate texture elements. The planar fit receiving support from the most regions is chosen as the correct interpretation. Simultaneously, those candidate texture elements that support the best plane are identified as the true texture elements.