Low level and intermediate level vision in aerial images

Low-level and intermediate-level computer vision tasks are regarded as transformations from lower to higher-level representations of the image information. An edge-based representation that makes explicit linear features and their spatial relationships is developed. Examples are presented in the scene domain of aerial images of urban scenes containing man-made structures.

The techniques used are based on a common structural and statistical model of the image data. This model assumes that the image data is adequately represented locally by a bivariate cubic polynomial plus additive independent Gaussian noise. This model, although simple, is shown to be useful for the design of effective computer vision solving tasks.

Four low-level computer vision modules are developed. First, a gradient operator which reduces sharply the gradient direction estimate bias that plagues current operators while also reducing sensitivity to noise. Secondly, a Bayes decision procedure for automatic gradient threshold selection that produces results which are superior to those obtained by the best subjective threshold. Thirdly, the new gradient operator and automatic gradient threshold selection are used in Haralick's directional zero-crossing edge operator resulting in improved performance. Finally, a graytone corner detector with significantly better probability of correct corner assignment than other corner detectors available in the literature.

Intermediate-level modules are developed for the construction of a number of intermediate level units from linear features. Among these is a linear segment extraction method that uses both, zero-crossing positional and angular information together with their distributional characteristics to accomplish optimal linear segment fitting. Methods for hypothesizing comers and relations of parallelism and collinearity among pairs of linear segments are developed. These relations are used to build higher-level groupings of linear segments that are likely to correspond to cultural objects.