Unsupervised Localization of Shapes using Statistical Models

F. Destrempes and M. Mignotte (Canada)



In this paper, we present a coherent Markov model for deformations of shapes based on the statistical distribution of the gradient vector field of the gray level and a proba bilistic model for reduction of dimension. This allows us to infer the posterior distribution of deformations given the data. The localization of a shape in an image is then vie wed as the minimizing of the corresponding Gibbs field. We use a stochastic optimization algorithm in order to find the optimal deformation, firstly with a simpler Gibbs field and then on the complete Gibbs field. This yields an unsu pervised method for localization of shapes, provided that the image is not prohibitively too complex.

Important Links:

Go Back