Learning Control Parameters of a Vision Process Using Contextual Information

Abstract

Two of the problems that the user of an image understanding system must continuously face are the choice of an appropriate algorithm and the setting of its associated parameters. These requirements mean that the user must have a fairly high degree of expertise with the algorithms to accomplish a given task effectively. If, on the other hand, the system itself is able to learn how to select among its algorithms and to set their parameters through its experience with similar tasks, it should be possible to reduce the need for operator expertise while improving efficiency at the same time. This paper presents a method to accomplish this goal. Contextual information computed from the task, and the input data is used to search for similar situations and determine whether or not an algorithm is applicable, and which parameters are suitable for it. Different approaches have been investigated as the basis for finding similar situations. The first one uses a measure of similarity between context, element values. The second one uses a categorization method based on conceptual clustering. The main problem is the need to deal with both numerical and categorical variables.

To demonstrate the efficiency of our approach, we describe experiments involving the use of a snake algorithm to perform the task of curvilinear feature extraction. Our implementation allows the various parameters of this technique to be context specific. We show in this setting how our system makes the use of a vision process, easier by reducing the needed, user, expertise and improving efficiency in obtaining the desired results.