Genetic adaptation of segmentation parameters

This work presents a method for the automatic adaptation of segmentation parameters based on Genetic Algorithms. An intuitive and computationally simple fitness function, which expresses the similarity between the segmentation result and a reference provi

  • PDF / 717,636 Bytes
  • 17 Pages / 439.37 x 666.142 pts Page_size
  • 98 Downloads / 180 Views

DOWNLOAD

REPORT


G. A. O. P. Costa1, R. Q. Feitosa1, 3, T. B. Cazes1, B. Feijó2 1

Department of Electrical Engineering, Catholic University of Rio de Janeiro (PUC-Rio), Brasil, (gilson, raul, tcazes)@ele.puc-rio.br

2

Department of Informatics, Catholic University of Rio de Janeiro (PUCRio), Brasil, [email protected]

3

Department of Computer Engineering, Rio de Janeiro State University (UERJ), Brasil

KEYWORDS: Object-based classification, optimization, genetic algorithm ABSTRACT: This work presents a method for the automatic adaptation of segmentation parameters based on Genetic Algorithms. An intuitive and computationally simple fitness function, which expresses the similarity between the segmentation result and a reference provided by the user, is proposed. The method searches the solution space for a set of parameter values that minimizes the fitness function. A prototype including an implementation of a widely used segmentation algorithm was developed to assess the performance of the method. A set of experiments with medium and high spatial resolution remote sensing image data was carried out and the method was able to come close to the ideal solutions.

680

G. A. O. P. Costa, R. Q. Feitosa, T. B. Cazes, B. Feijó

1 Introduction The key step in object-based image interpretation is the segmentation of the image (Blaschke and Strobl 2001). In fact, the performance of the whole interpretation strongly depends on the segmentation quality. Therefore, proper segmentation parameters must be chosen before starting the classification process. The relation between the parameter values and the corresponding segmentation outcome, however, is in most cases far from being obvious, and the definition of suitable parameter values is usually done through a troublesome and time consuming trial and error process. Many semiautomatic approaches have been proposed to reduce the burden of parameter adaptation, starting from simple graphic support tools (e.g. Schneider et al. 1997), going through interactive systems (e.g. Matsuyama 1993) in which the user is asked to rate the result after each adaptation iteration (Crevier and Lepage 1997), up to nearly automatic solutions that require a minimum of human intervention. The automatic adaptation of segmentation parameters involves two main issues: the selection of an objective function that expresses adequately the quality of the segmentation (Bhanu et al. 1995); and the choice of the optimization method for the search of parameter values that maximize the objective function. In supervised methods the quality measure reflects the similarity among the segmentation output and reference segments usually produced manually by a photo-interpreter (Zhang 1996). Unsupervised methods, on the contrary, use no references and do not consider human induced subjectivity or application particularities (Espindola et al. 2006). Generally the relationship among the segmentation parameter values and the quality measure can not be formulated analytically. In such cases calculus based optimization methods cannot be